The way I see it, artificial intelligence (or AI), really leapt into the zeitgeist in late-2022 or early-2023 with the public introduction of DALL-E2 and ChatGPT. Both are provided by OpenAI and are software products that use AI to generate art and writing, respectively (and often at astounding quality). Since then, developments in AI have progressed at a breathtaking pace.

With the latest earnings season for the US stock market – for the second quarter of 2024 – coming to its tail-end, I thought it would be useful to collate some of the interesting commentary I’ve come across in earnings conference calls, from the leaders of technology companies that I follow or have a vested interest in, on the topic of AI and how the technology could impact their industry and the business world writ large. This is an ongoing series. For the older commentary:

• 2023 Q1 – here and here
• 2023 Q2 – here and here
• 2023 Q3 – here and here
• 2023 Q4 – here and here
• 2024 Q1 – here and here

With that, here are the latest commentary, in no particular order:

Airbnb (NASDAQ: ABNB)

Airbnb’s management is still really excited about AI, but they’ve also realised that it’s going to take a lot longer for applications to change; management sees three layers to AI, namely, the chip, the model, and the application, and while there’s been a lot of innovation on the chip and the model, not much has changed with the applications, especially in e-commerce and travel

ChatGPT launched late November 2022. When it launched, I think we all got like incredibly excited. It was kind of like the moment probably some of us first discovered the Internet or maybe when iPhone was launched. And when it was launched, you had the feeling that everything was going to change. But I think that’s still true. But I think one of the things we’ve learned over the last, say, 18 months or nearly 2 years — 22 months since ChatGPT launched is that it’s going to take a lot longer than people think for applications to change.

If I were to think of AI, I’d probably think about it in 3 layers. You have the chip. You have the model. And you have the applications. There’s been a lot of innovation on the chip. There’s been a lot of innovation on the model. We have a lot of new models, and there’s a prolific rate of improvement in these models. But if you look at your home screen, which of your apps are fundamentally different because of the AI, like fundamentally different because of generative AI? Very little, especially even less in e-commerce or travel. And the reason why is I think it’s just going to take time to develop new AI paradigm. 

Airbnb’s management sees ChatGPT, even though it’s an AI chat software, as an application that could have existed before AI; management thinks what needs to be done is to develop AI applications that are native to the AI models with unique interfaces and no one has done this year; Airbnb is working on an application that will be native to AI models and this will change how users interact with Airbnb, where it becomes much more than a search box; this change in Airbnb will take a few years to develop

ChatGPT [ is an AI model interface that could ] have existed before AI. And so all of our paradigms are pre-AI paradigms. And so what we need to do is we need to actually develop AI applications that are native to the model. No one has done this yet. There’s not been one app that I’m aware of that’s the top 50 app in the app store in the United States that is a fundamentally new paradigm as fundamentally different as a multitouch was to the iPhone in 2008, and we need that interface change. So that’s one of the things that we’re working on. And I do think Airbnb will eventually be much more than a search box where you type a destination, add dates and find a listing. It’s going to be much more of a travel concierge. It’s having a conversation, learning, adapting to you. It’s going to take a number of years to develop this. And so it won’t be in the next year that this will happen. And I think this is probably what most of my tech friends are also saying, is it’s going to just take a bit more time.

Airbnb’s management thinks that having a new AI-driven interface will allow Airbnb to expand into new businesses

But to answer your question on what’s possible, a new interface paradigm would allow us to attach new businesses. So the question is, what permission do we have to go into a business like hotels? Well, today, we have permission because we have a lot of traffic. But if we had a breakthrough interface, we have even more permission because suddenly, we could move top of funnel and not just ask where are you going, but we can point to — we can inspire where you travel. Imagine if we had an index of the world’s communities. We told you we had information about every community, and we can provide the end-to-end trip for you. So there’s a lot of opportunities as we develop new interfaces to cross-sell new more inventory. 

Alphabet (NASDAQ: GOOG)

Google Cloud’s year-to-date AI-related revenue is already in the billions, and its AI infrastructure and solutions are already used by >2 million developers; more than 1.5 million developers are using Gemini, Alphabet’s foundational AI model, across the company’s developers tools

Year-to-date, our AI infrastructure and generative AI solutions for cloud customers have already generated billions in revenues and are being used by more than 2 million developers…

…More than 1.5 million developers are now using Gemini across our developer tools.

Alphabet’s management thinks Alphabet is well-positioned for AI; Alphabet is innovating at every layer of the AI stack, from chips at the bottom to agents at the top

As I spoke about last quarter, we are uniquely well positioned for the AI opportunity ahead. Our research and infrastructure leadership means we can pursue an in-house strategy that enables our product teams to move quickly. Combined with our model building expertise, we are in a strong position to control our destiny as the technology continues to evolve. Importantly, we are innovating at every layer of the AI stack, from chips to agents and beyond, a huge strength.

Alphabet’s management thinks Alphabet is using AI to deliver better responses on Search queries; tests for AI Overviews has showed increase in Search usage and higher user satisfaction; Search users with complex searches keep coming back for AI Overviews; users aged 18-24 have higher engagement when using Search with AI Overviews; Alphabet is prioritising AI-approaches that send traffic to websites; ads that are above or below AI Overviews continue to be valuation; in 2024 Q2, management has doubled the core model size for AI Overviews while improving latency and keeping cost per AI Overviews served flat; management is working on matching the right AI model size to the query’s complexity to improve cost and latency; AI Overviews is rolled out in the USA and will be rolled out to more countries throughout 2024; Alphabet will soon put Search and Shopping ads within the AI Overviews for USA users

With AI, we are delivering better responses on more types of search queries and introducing new ways to search. We are pleased to see the positive trends from our testing continue as we roll out AI Overviews, including increases in Search usage and increased user satisfaction with the results. People who are looking for help with complex topics are engaging more and keep coming back for AI Overviews. And we see even higher engagement from younger users aged 18 to 24 when they use Search with AI Overviews. As we have said, we are continuing to prioritize approaches that send traffic to sites across the web. And we are seeing that ads appearing either above or below AI Overviews continue to provide valuable options for people to take action and connect with businesses…

…Over the past quarter, we have made quality improvements that include doubling the core model size for AI Overviews while at the same time improving latency and keeping cost per AI Overviews served flat. And we are focused on matching the right model size to the complexity of the query in order to minimize impact on cost and latency…

…On the AI Overviews, we are — we have rolled it out in the U.S. And we are — will be, through the course of the year, definitely scaling it up, both to more countries…

…And as you have probably noticed at GML, we announced that soon we’ll actually start testing search and shopping ads in AI Overviews for users in the U.S., and they will have the opportunity to actually appear within the AI Overview in a section clearly labeled as sponsored when they’re relevant to both the quarry and the information in the AI Overview, really giving us the ability to innovate here and take this to the next level.

AI opens up new ways to use Search, such as asking questions by taking a video with Lens; AI Overviews in Lens has led to higher overall visual search usage; Circle to Search is another new way to search, and is available on >100 million Android devices

AI expands the types of queries we are able to address and opens a powerful new ways to Search. Visual search via Lens is one. Soon, you’ll be able to ask questions by taking a video with Lens. And already, we have seen that AI Overviews in Lens leads to an increase in overall visual search usage. Another example is Circle to Search, which is available today on more than 100 million Android devices.

Gemini, now has 4 sizes, each with their own use cases; Gemini comes with a context window of 2 million, the longest of any foundation model to-date; all of Alphabet’s 6 products with more than 2 billion monthly users are using Gemini; through Gemini, users of Google Photos can soon ask questions of their photos and receive answers

Gemini now comes in 4 sizes with each model designed for its own set of use cases. It’s a versatile model family that runs efficiently on everything from data centers to devices. At 2 million tokens, we offer the longest context window of any large-scale foundation model to date, which powers developer use cases that no other model can handle. Gemini is making Google’s own products better. All 6 of our products with more than 2 billion monthly users now use Gemini…

…At I/O, we showed new features coming soon to Gmail and to Google Photos. Soon, you’ll be able to ask Photos questions like, what did I eat at that restaurant in Paris last year?

During Alphabet’s recent developer conference, I/O, management showed their vision of what a universal AI agent could look like

For a glimpse of the future, I hope you saw Project Astra at I/O. It shows multimodal understanding and natural conversational capabilities. We’ve always wanted to build a universal agent, and it’s an early look at how they can be helpful in daily life.

Alphabet has launched Trillium, the sixth-generation of its custom TPU AI accelerator; Trillium has a 5x increase in peak compute performance per chip and a 67% improvement in energy efficiency over TPU v5e

Trillium is the sixth generation of our custom AI accelerator, and it’s our best-performing and most energy-efficient TPU to date. It achieves a near 5x increase in peak compute performance per chip and a 67% more energy efficient compared to TPU v5e.

Google Cloud’s enterprise AI platform, Vertex, is used by Deutsche Bank, Kingfisher, and the US Air Force to build AI agents; Uber and WPP are using Gemini Pro 1.5 and Gemini Flash 1.5 in Vertex for customer experience and marketing; Vertex has broadened support for 3rd-party AI models, including Anthropic’s Claude 3.5 Sonnet, Meta’s Llama, and Mistral’s models

Our enterprise AI platform, Vertex, helps customers such as Deutsche Bank, Kingfisher and the U.S. Air Force build powerful AI agents. Last month, we announced a number of new advances. Uber and WPP are using Gemini Pro 1.5 and Gemini Flash 1.5 in areas like customer experience and marketing. We broadened support for third-party models including Anthropic’s Claude 3.5 Sonnet and open-source models like Gemma 2, Llama and Mistral. 

Google Cloud is the only cloud provider to provide grounding with Google Search; large enterprises such as Moody’s, MSCI, and ZoomInfo are using Google Cloud’s grounding capabilities

We are the only cloud provider to offer grounding with Google Search, and we are expanding grounding capabilities with Moody’s, MSCI, ZoomInfo and more.

Google Cloud’s AI-powered applications are helping it to drive upsells and win new customers; Best Buy and Gordon Food Service are using Google Cloud’s conversational AI platform; Click Therapeutics is using Gemini for Workspace; Wipro is using Gemini Code Assist to speed up software development; MercadoLibre is using BigQuery and Looker for capacity planning and speeding up shipments.

Our AI-powered applications portfolio is helping us win new customers and drive upsell. For example, our conversational AI platform is helping customers like Best Buy and Gordon Food Service. Gemini for Workspace helps Click Therapeutics analyze patient feedback as they build targeted digital treatments. Our AI-powered agents are also helping customers develop better-quality software, find insights from their data and protect their organization against cybersecurity threats using Gemini. Software engineers at Wipro are using Gemini Code Assist to develop, test and document software faster. And data analysts at Mercado Libre are using BigQuery and Looker to optimize capacity planning and fulfill shipments faster.

 In 2024 Q2, Alphabet announced more than 30 new ads features and products to help advertisers leverage AI; Alphabet is applying AI across its advertising products to streamline workflows, enhance asset creation, and improve engagement with consumers; in asset creation, any business using Product Studio can upload an image and enhance it with AI; AI features for consumers such as virtual try-ons in shopping ads are in beta-testing, and feedback shows that virtual try-on gets 60% more high-quality views; advertisers using Alphabet’s AI-powered profit maximisation tools along with Smart Bidding see a 15% increase in profit; Demand Gen, to be rolled out in the coming months, creates high-quality image assets for social marketers and delivers 14% more conversions when paired with Search or Performance Max; Tiffany used Demand Gen and achieved a 2.5% lift in consideration and important customer-actions, and a 5.6x improvement in cost per click compared to social media benchmarks; Alphabet used Demand Gen to create 4,500 ad variations for Pixel 8’s advertising campaigns and delivered twice the clicks per rate at nearly 1/4 of the cost

This quarter, we announced over 30 new ads features and products to help advertisers leverage AI and keep pace with the evolving expectations of customers and users. Across Search, PMax, Demand Gen and retail, we’re applying AI to streamline workflows, enhance creative asset production and provide more engaging experiences for consumers.

Listening to our customers, retailers in particular have welcomed AI-powered features to help scale the depth and breadth of their assets. For example, as part of the new and easier-to-use Merchant Center, we’ve expanded Product Studio with tools that bring the power of Google AI to every business owner. You can upload a product image, prompt the AI with something like feature this product with Paris skyline in the background, and Product Studio will generate campaign-ready assets.

I also hear great feedback from our customers on many of our other new AI-powered features. We’re beta testing virtual try-on in shopping ads and plan to roll it out widely later this year. Feedback shows this feature gets 60% more high-quality views than other images and higher click out to retailer sites. Retailers love it because it drives purchasing decisions and fewer returns.

Our AI-driven profit optimization tools have been expanded to Performance Max and standard shopping campaigns. Advertisers who use profit optimization and Smart Bidding see a 15% uplift in profit on average compared to revenue-only bidding.

Lastly, Demand Gen is rolling out to Display & Video 360 and Search Ads 360 in the coming months with new generative image tools that create stunning, high-quality image assets for social marketers. As we said at GML, when paired with Search or PMax, Demand Gen delivers an average of 14% more conversions…

…Luxury jewelry retailer Tiffany leveraged Demand Gen during the holiday season and saw a 2.5% brand lift in consideration and actions such as adding items to carts and booking appointments. The campaign drove a 5.6x more efficient cost per click compared to social media benchmarks. Our own Google marketing team used Demand Gen to create nearly 4,500 ad variations for Pixel 8 campaign shown across YouTube, Discover and Gmail, delivering twice the clicks per rate at nearly 1/4 of the cost.

Alphabet has used AI to (1) improve broad match performance by 10% in 6 months for advertisers using Smart Bidding, and (2) increase conversions by 25% at similar cost for advertisers who adopt PMax to broad match and Smart Bidding in their Search campaigns

In just 6 months, AI-driven improvements to quality, relevance and language understanding have improved broad match performance by 10% for advertisers using Smart Bidding. Also, advertisers who adopt PMax to broad match and Smart Bidding in their Search campaigns see an average increase of over 25% more conversions of value at a similar cost.

Google Cloud had 29% revenue growth in 2024 Q2 (was 28% in 2024 Q1); operating margin was 11% (was 9% in 2024 Q1 and was 4.9% in 2023 Q2); Google Cloud’s accelerating revenue growth in 2024 Q2 was partly the result of AI demand; GCP’s growth rate is above the growth rate for the overall Google Cloud business

Turning to the Google Cloud segment. Revenues were $10.3 billion for the quarter, up 29%, reflecting, first, significant growth in GCP, which was above growth for Cloud overall and includes an increasing contribution from AI; and second, strong Google Workspace growth, primarily driven by increases in average revenue per seat. Google Cloud delivered operating income of $1.2 billion and an operating margin of 11%…

…[Question] On the cloud acceleration, would you characterize that as new AI demand helping drive that year-to-date? Or is that more of a rebound in just general compute and other demand?

[Answer] There is clearly a benefit as the Cloud team is engaging broadly with customers around the globe with AI-related solutions, AI infrastructure solutions and generative AI solutions. I think we noted that we’re particularly encouraged that the majority of our top 100 customers are already using our generative AI solution. So it is clearly adding to the strength of the business on top of all that they’re doing. And just to be really clear, the results for GCP, the growth rate for GCP is above the growth for Cloud overall.

Alphabet’s big jump capex in 2024 Q2 (was $7.2 billion in 2023 Q2) was mostly for technical infrastructure, in the form of servers and data centers; management continues to expect Alphabet’s quarterly capex for the rest of 2024 to be similar to what was seen in 2024 Q1;

With respect to CapEx, our reported CapEx in the second quarter was $13 billion, once again, driven overwhelmingly by investment in our technical infrastructure with the largest component for servers followed by data centers. Looking ahead, we continue to expect quarterly CapEx throughout the year to be roughly at or above the Q1 CapEx of $12 billion, keeping in mind that the timing of cash payments can cause variability in quarterly reported CapEx.

Alphabet’s management is seeing more tangible use cases for AI in the consumer space compared to the enterprise space; in the consumer space, consumers are engaging with Alphabet’s AI features, but there’s still the question of monetisation; in the enterprise space, a lot of AI models are currently being built and they are converging towards a set of base capabilities; the next wave for the enterprise space will be building applications on top of the models, and there is some traction in some areas, but it’s not widespread yet; management believes value will eventually be unlocked, but it may take time

 I think there is a time curve in terms of taking the underlying technology and translating it into meaningful solutions across the board, both on the consumer and the enterprise side. Definitely, on the consumer side, I’m pleased, as I said in my comments earlier, in terms of how for a product like Search, which is used at that scale over many decades, how we’ve been able to introduce it in a way that it’s additive and enhances overall experience and this positively contributing there. I think across our consumer products, we’ve been able — I think we are seeing progress on the organic side. Obviously, monetization is something that we would have to earn on top of it. The enterprise side, I think we are at a stage where definitely there are a lot of models. I think roughly, the models are all kind of converging towards a set of base capabilities. But I think where the next wave is working to build solutions on top of it. And I think there are pockets, be it coding, be it in customer service, et cetera, where we are seeing some of those use cases are seeing traction, but I still think there is hard work there to completely unlock those…

…But I think we are in this phase where we have to deeply work and make sure on these use cases, on these workflows, we are driving deeper progress on unlocking value, which I’m very bullish will happen. But these things take time. So — but if I were to take a longer-term outlook, I definitely see a big opportunity here. And I think particularly for us, given the extent to which we are investing in AI, our research infrastructure leadership, all of that translates directly. And so I’m pretty excited about the opportunity space ahead.

Alphabet’s management thinks that the risk of underinvesting in AI infrastructure for the cloud business is currently greater than the risk of overinvesting; management thinks that even if Alphabet ends up overinvesting, the infrastructure is still widely useful for internal use cases

[Question] So it looks like from the outside at least, the hyperscaler industry is going from kind of an underbuilt situation this time last year to better meeting the demand with capacity right now to potentially being overbuilt next year if these CapEx growth rates keep up. So do you think that’s a fair characterization? And how are we thinking about the return on invested capital with this AI CapEx cycle?

[Answer] I think the one way I think about it is when we go through a curve like this, the risk of under-investing is dramatically greater than the risk of over-investing for us here, even in scenarios where if it turns out that we are over-investing, we clearly — these are infrastructure which are widely useful for us. They have long useful lives, and we can apply it across, and we can work through that. But I think not investing to be at the front here, I think, definitely has much more significant downside. Having said that, we obsess around every dollar we put in. Our teams are — work super hard. I’m proud of the efficiency work, be it optimization of hardware, software, model deployment across our fleet. All of that is something we spend a lot of time on, and that’s how we think about it.

Amazon (NASDAQ: AMZN)

AWS’s AI business continues to grow dramatically with a multi-billion revenue run rate; management sees AWS’s AI services resonating with customers, who want choice in the AI models and AI chips they use, and AWS is providing them with choices; over the past 18 months, AWS has launched twice as many AI features into general availability than all other major cloud providers combined

Our AI business continues to grow dramatically with a multibillion-dollar revenue run rate despite it being such early days, but we can see in our results and conversations with customers that our unique approach and offerings are resonating with customers. At the heart of this strategy is a firmly held belief, which we’ve had since the beginning of AWS that there is not one tool to rule the world. People don’t want just one database option or one analytics choice or one container type. Developers and companies not only reject it but are suspicious of it. They want multiple options for flexibility and to use the best tool for each job to be done. The same is true in AI. You saw this several years ago when some companies tried to argue that TensorFlow will be the only machine learning framework that mattered and then PyTorch and others overtook it. The same one model or one chip approach dominated the earliest moments of the generative AI boom, but we have a lot of data that suggests this is not what customers want here either, and our AWS team is determined to deliver choice and options for customers…

…During the past 18 months, AWS has launched more than twice as many machine learning and generative AI features into general availability than all the other major cloud providers combined. 

AWS provides NVIDIA chips for AI model builders, but management also hear from customers that they want better price performance and hence AWS developed the Trainium and Inferentia chips for training and inference, respectively; the second version of Trainium is coming later this year and has very compelling price performance; management is seeing significant demand for Trainium and Inferentia; management started building Trainium and Inferentia 5 years ago also because they had the experience of seeing customers wanting better price performance from CPUs; management believes Trainium and Inferentia will generate similarly high ROI as Graviton, Amazon’s custom CPU, does

For those building generative AI models themselves, the cost of compute for training and inference is critical, especially as models get to scale. We have a deep partnership with NVIDIA and the broader selection of NVIDIA instances available, but we’ve heard loud and clear from customers that they relish better price performance. It’s why we’ve invested in our own custom silicon in Trainium for training and Inferentia for inference. And the second versions of those chips, with Trainium coming later this year, are very compelling on price performance. We are seeing significant demand for these chips…

…When we started AWS, we had and still have a very deep partnership with Intel on the generalized CPU space. But what we found from customers is that they — when you find a — an offering that is really high value for you and high return, you don’t actually spend less, even though you’re spending less per unit. You spend less per unit, but it enables you, it frees you up to do so much more inventing and building for your customers. And then when you’re spending more, you actually want better price performance than what you’re getting.

And a lot of times, it’s hard to get that price performance from existing players unless you decide to optimize yourself for what you’re learning from your customers and you push that envelope yourself. And so we built custom silicon in the generalized CPU space with Graviton, which we’re on our fourth model right now. And that has been very successful for customers and for our AWS business, is it saves customers about — up to about 30% to 40% price performance versus the other leading x86 processors that they could use.

And we saw the same trend happening about 5 years ago in the accelerator space in the GPU space, where the products are good, but there was really primarily 1 provider and supply was more scarce than what people wanted. And people — our customers really want improved price performance all the time. And so that’s why we went about building Trainium, which is our training chip, and Inferentia, which is our inference chip, which we’re on second versions of both of those. They will have very compelling relative price performance.

And in a world where it’s hard to get GPUs today, the supply is scarce and all the schedules continue to move over time, customers are quite excited and demanding at a high clip, our custom silicon, and we’re producing it as fast as we can. I think that’s going to have very good return profile just like Graviton has, and I think it will be another differentiating feature around AWS relative to others.

SageMaker, AWS’s fully-managed AI service, helps customers save time and money while they build their AI models; management is seeing model builders standardise on SageMaker

Model builders also desire services that make it much easier to manage the data, construct the models, experiment, deploy to production and achieve high-quality performance, all while saving considerable time and money. That’s what Amazon SageMaker does so well including its most recently launched feature called HyperPod that changes the game and networking performance for large models, and we’re increasingly seeing model builders standardize on SageMaker. 

Amazon Bedrock, AWS’s AI-models-as-a-service offering, caters to companies that want to leverage 3rd-party models and customise with their own data; Bedrock already has tens of thousands of companies using it; Bedrock has the largest selection of models and the best generative AI capabilities in a number of critical areas; Bedrock recently added Anthropic’s Claude 3.5 models, Meta’s new Llama 3.1 models, and Mistral’s new models

While many teams will build their own models, lots of others will leverage somebody else’s frontier model, customize it with their own data, and seek a service that provides broad model selection and great generative AI capabilities. This is what we think of as the middle layer, what Amazon Bedrock does and why Bedrock has tens of thousands of companies using it already. Bedrock has the largest selection of models, the best generative AI capabilities in critical areas like model evaluation, guardrails, RAG and agenting and then makes it easy to switch between different model types and model sizes. Bedrock has recently added Anthropic’s Claude 3.5 models, which are the best performing models on the planet; Meta’s new Llama 3.1 models; and Mistral’s new Large 2 models. And Llama’s and Mistral’s impressive performance benchmarks and open nature are quite compelling to our customers as well.

Amazon’s management is seeing strong adoption of Amazon Q, Amazon’s generative AI assistant for software development; Amazon Q has the highest score and acceptance rate for code suggestions; Amazon Q tests code and outperforms competitors on catching security vulnerabilities; with Amazon Q’s code transformation capabilities, Amazon saved $260 million and 4,500 developer years when performing a large Java Development Kit migration; management thinks Amazon Q can continue to improve and address more use cases  

We’re continuing to see strong adoption of Amazon Q, the most capable generative AI-powered assistant for software development and to leverage your own data. Q has the highest known score and acceptance rate for code suggestions, but it does a lot more than provide code suggestions. It tests code, outperforms all other publicly benchmarkable competitors on catching security vulnerabilities and leads all software development assistance on connecting multiple steps together and applying automatic action.

It also saves development teams time and money on the muck nobody likes to talk about. For instance, when companies decide to upgrade from one version of a framework to another, it takes development teams many months, sometimes years burning valuable opportunity costs and churning developers who hate this tedious though important work. With Q’s code transformation capabilities, Amazon has migrated over 30,000 Java JDK applications in a few months, saving the company $260 million and 4,500 developer years compared to what it would have otherwise cost. That’s a game changer.

And think about how this Q transformation capability might evolve to address other elusive but highly desired migrations. 

Amazon’s management is still very bullish on the medium to long-term impacts of AI, but the progress may not be a straight line; management sees a lot of promise in generative AI being able to improve customer experiences and this is informed by their own experience of using generative AI within Amazon, such as: (1) Rufus, a shopping assistant, improves customers’ shopping decisions, (2) customers can virtually try on apparel, (3) sellers can create new selections with a line or two of text, and (4) better detection of product defects before the products reach customers

We remain very bullish on the medium to long-term impact of AI in every business we know and can imagine. The progress may not be one straight line for companies.

Generative AI especially is quite iterative, and companies have to build muscle around the best way to solve actual customer problems. But we see so much potential to change customer experiences. We see it in how our generative-AI-powered shopping assistant, Rufus, is helping customers make better shopping decisions. We see it in our AI features that allow customers to simulate trying apparel items or changing the buying experience. We see it in our generative AI listing tools enabling sellers to create new selection with a line or 2 of text versus the many forms previously required. We see it in our fulfillment centers across North America, where we’re rolling out Project Private Investigator, which uses a combination of generative AI and computer vision to uncover defects before products reach customers. We see it in how our generative AI is helping our customers discover new music and video. We see it in how it’s making Alexa smart, and we see it in how our custom silicon and services like SageMaker and Bedrock are helping both our internal teams and many thousands of external companies reinvent their customer experiences and businesses. We are investing a lot across the board in AI, and we’ll keep doing so as we like what we’re seeing and what we see ahead of us.

Amazon’s management expects capital expenditures to be higher in 2024 H2 compared to 2024 H1; most of the capex will be for AWS infrastructure in both generative AI and non-generative AI workloads; management has a lot of experience, accumulated over the years, in predicting just the right amount of compute capacity to provide for AWS before the generative AI era, and they believe they can do so again for generative AI; management is investing heavily in AI-related capex because they see a lot of demand and in fact, they would like AWS to have more compute capacity than what it has today

For the first half of the year, CapEx was $30.5 billion. Looking ahead to the rest of 2024, we expect capital investments to be higher in the second half of the year. The majority of the spend will be to support the growing need for AWS infrastructure as we continue to see strong demand in both generative AI and our non-generative AI workloads…

…If you think about the fact that we have about 35 regions and think of a region as multiple — a cluster of multiple data centers and about 110 availability zones, which is roughly equivalent to a data center, sometimes it includes multiple and then if you think about having to land thousands and thousands of SKUs across the 200 AWS services in each of those availability zones at the right quantities, it’s quite difficult. And if you end up actually with too little capacity, then you have service disruptions, which really nobody does because it means companies can’t scale their applications.

So most companies deliver more capacity than they need. However, if you actually deliver too much capacity, the economics are pretty woeful, and you don’t like the returns of the operating income. And I think you can tell from having — we disclosed both our revenue and our operating income in AWS that we’ve learned over time to manage this reasonably well. And we have built models over a long period of time that are algorithmic and sophisticated that land the right amount of capacity. And we’ve done the same thing on the AI side.

Now AI is newer. And it’s true that people take down clumps of capacity in AI that are different sometimes. I mean — but it’s also true that it’s not like a company shows up to do a training cluster asking for a few hundred thousand chips the same day. Like you have a very significant advanced signal when you have customers that want to take down a lot of capacity.

So while the models are more fluid, it’s also true that we’ve built, I think, a lot of muscle and skill over time in building these capacity signals and models, and we also are getting a lot of signal from customers on what they need. I think that it’s — the reality right now is that while we’re investing a significant amount in the AI space and in infrastructure, we would like to have more capacity than we already have today. I mean we have a lot of demand right now, and I think it’s going to be a very, very large business for us.

Companies need to organise their data in specific ways before they can use AI effectively; it’s difficult for companies with on-premise data centers to use AI effectively

It’s quite difficult to be able to do AI effectively if your data is not organized in such a way that you can access that data and run the models on top of them and then build the application. So when we work with customers, and this is true both when we work directly with customers as well as when we work with systems integrator partners, everyone is in a hurry to get going on doing generative AI. And one of the first questions that we ask is show us where your data is, show us what your data lake looks like, show us how you’re going to access that data. And there’s very often work associated with getting your data in the right shape and in the right spot to be able to do generative AI. There — fortunately, because so many companies have done the work to move to the cloud, there’s a number of companies who are ready to take advantage of AI, and that’s where we’ve seen a lot of the growth. But also it’s worth remembering that, again, remember the 90% of the global IT spend being on-premises. There are a lot of companies who have yet to move to the cloud, who will, and the ability to use AI more effectively is going to be one of the many drivers in doing so for them.

Apple (NASDAQ: AAPL)

Apple Intelligence, Apple’s AI technologies embedded in its devices, improves Siri; Apple Intelligence is built on a foundation of privacy and has a ground-breaking approaching to using the cloud, known as Private Cloud Compute, that protects user information; Apple Intelligence is powered by Apple’s custom chips; Apple Intelligence will involve integration with ChatGPT in iPhones, Macs, and iPads; management will continue to invest in AI; because of management’s stance on privacy, Apple Intelligence will maximise the amount of data that is processed directly on people’s devices; Apple Intelligence’s roll out will be staggered; Apple Intelligence’s monetisation appears to involve both the Services business of Apple, and payments from partners

At our Worldwide Developers Conference, we were thrilled to unveil game-changing updates across our platforms, including Apple Intelligence. Apple Intelligence builds on years of innovation and investment in AI and machine learning. It will transform how users interact with technology from Writing Tools to help you express yourself to Image Playground, which gives you the ability to create fun images and communicate in new ways, to powerful tools for summarizing and prioritizing notifications. Siri also becomes more natural, more useful, and more personal than ever. Apple Intelligence is built on a foundation of privacy, both through on-device processing that does not collect users’ data and through Private Cloud Compute, a groundbreaking new approach to using the cloud while protecting users’ information powered by Apple Silicon. We are also integrating ChatGPT into experiences within iPhone, Mac, and iPad, enabling users to draw on a broad base of world knowledge.

We are very excited about Apple Intelligence, and we remain incredibly optimistic about the extraordinary possibilities of AI and its ability to enrich customers’ lives. We will continue to make significant investments in this technology and dedicate ourselves to the innovation that will unlock its full potential…

…We are committed as ever to shipping products that offer the highest standards of privacy for our users. With everything we do, whether it’s offering a browser like Safari that prevents third-parties from tracking you across the Internet, or providing new features like the ability to lock and hide apps, we are determined to keep our users in control of their own data. And we are just as dedicated to ensuring the security of our users’ data. That’s why we work to minimize the amount of data we collect and work to maximize how much is processed directly on people’s devices, a foundational principle that is at the core of all we build, including Apple Intelligence…

…The rollout, as we mentioned in June, sort of we’ve actually started with developers this week. We started with some features of Apple Intelligence, not the complete suite. There are other features like languages beyond U.S. English that will happen over the course of the year, and there are other features that will happen over the course of the year. And ChatGPT is integrated by the end of the calendar year. And so yes, so it is a staggered launch…

…[Question] How should investors think about the monetization models…  in the long term, do you see the Apple Intelligence part, the Services growth from Apple Intelligence being the larger contributor over time? Or do you see these partnerships becoming a larger contributor over time? 

[Answer] The monetization model, I don’t want to get into the terms of the commercial agreements because they’re confidential between the parties, but I see both aspects as being very important. People want both.

Apple is getting its partners to fork out the bill for some of its capex needs for AI cloud compute, so even though its capex will increase over time, it does not seem like the increase may be that high

[Question] Do you see the rollout of these features requiring further increases in R&D or increases in OpEx or CapEx for cloud compute capacity?

[Answer] On the CapEx part, it’s important to remember that we employ a hybrid kind of approach where we do things internally and we have certain partners that we do business with externally where the CapEx would appear in their respective businesses. But yes, I mean, you can expect that we will continue to invest and increase it year-on-year…

…On the CapEx front, as Tim said, we employ a hybrid model. Some of the investments show up on our balance sheet and some other investments show up somewhere else and we pay as we go. But in general, we try to run the company efficiently.

Arista Networks (NYSE: ANET)

Arista Networks recently launched its Etherlink AI platforms that are compatible with the ultra-Ethernet consortium and can lead the migration from Infiniband to Ethernet; the Etherlink AI platforms consist of a portfolio of 800-gig switches and can work with all kinds of GPUs; there are new products in the platform that work well even for very large AI clusters; the Etherlink portfolio is being trialled by customers can support up to be 100,000 XPUs

In June 2024, we launched Arista’s Etherlink AI platforms that are ultra-Ethernet consortium compatible, validating the migration from InfiniBand to Ethernet. This is a rich portfolio of 800-gig products, not just a point product, but in fact, a complete portfolio that is both NIC and GPU agnostic. The AI portfolio consists of the 7060 [indiscernible] switch that supports 64 800-gig or 128 400-gig Ethernet ports with a capacity of 51 terabits per second. The 7800 R4 AI Spine is our fourth generation of Arista’s flagship 7800, offering 100% non-blocking throughput with a proven virtual output queuing architecture. The 7800 R4 supports up to 460 terabits in a single chassis, corresponding to 576800 gigabit Ethernet ports or 1,152400 gigabit port density. The 7700 R4 AI distributed Etherlink Switch is a unique product offering with a massively parallel distributed scheduling and congestion-free traffic spraying fabric. The 7700 represents the first in a new series of ultra-scalable intelligent distributed systems that can deliver the highest consistent throughput for very large AI clusters…

…Our Etherlink portfolio is in the midst of trials and can support up to 100,000 XPUs in a 2-tier design built on our proven and differentiated extensible OS.

Arista Networks had a recent AI enterprise win with a Tier 2 cloud provider to provide Ethernet fabrics for its fleet of NVIDIA H100 GPUs; the cloud provider was using a legacy networking vendor that could not scale

The first example is an AI enterprise win with a large Tier 2 cloud provider which has been heavily investing in GPUs to increase their revenue and penetrate new markets. Their senior leadership wanted to be less reliant on traditional core services and work with Arista on new, reliable and scalable Ethernet fabrics. Their environment consisted of new NVIDIA H100s. However, it was being connected to their legacy networking vendor, which resulted in them having significant performance and scale issues with their AI applications. The goal of our customer engagement was to refresh the front-end network to alleviate these issues. Our technical partnership resulted in deploying a 2-step migration path to alleviate the current issues using 400-gig 7080s, eventually migrating them to an 800-gig AI Ethernet link in the future. 

Arista Networks’ management is once again seeing the network becoming the computer as AI training models require a lossless network to connect every AI accelerator in a cluster to one another; Arista Networks’ AI networking solutions also connect trained AI models to end users and other systems

I am reminded of the 1980s when Sun [Microsystems] for declaring the network is the computer. Well, 40 years later, we’re seeing the same cycle come true again with the collective nature of AI training models mandating a lossless highly available network to seamlessly connect every AI accelerator in the cluster to one another for peak job completion times. Our AI networks also connect trained models to end users and other multi-tenant systems in the front-end data center, such as storage, enabling the AI system to become more than the sum of its parts.

Arista Networks’ management think that data centers will evolve to be holistic AI centers, where the network will be the epicenter; management thinks that AI centers will need a foundational data architecture; Arista Networks has an AI agent within its EOS (Extensible Operating System) that can connect to NVIDIA’s Bluefield NICs (network interface cards), with more NICs to be added in the future

We believe data centers are evolving to holistic AI centers, where the network is the epicenter of AI management for acceleration of applications, compute, storage and the wide area network. AI centers need a foundational data architecture to deal with the multimodal AI data sets that run on our differentiated EOS network data systems. Arista showcased the technology demonstration of our EOS-based AI agent that can directly connect on the NIC itself or alternatively, inside the host. By connecting into adjacent Arista switches to continuously keep up with the current state, send telemetry or receive configuration updates, we have demonstrated the network working holistically with network interface cards such as NVIDIA Bluefield and we expect to add more NICs in the future.

Arista Networks’ management thinks that as GPUs increase in speed, the dependency on the network for higher throughput increases

I think as the GPUs get faster and faster, obviously, the dependency on the network for higher throughput is clearly related.

The 4 major AI trials Arista Networks discussed in the 2024 Q1 earnings call are all going well and ar removing into pilots these year

[Question] Last quarter, you had mentioned kind of 4 major AI trials that you guys were a part of…  any update on where those 4 AI trials stand or what the current count of AI trials is currently?

[Answer] All 4 trials are largely in what I call Cloud and AI Titans. A couple of them could be classified as specialty providers as well, depending on how they end up. But those 4 are going very well. They started out as largely trials. They’re now moving into pilots this year, most of them. 

Arista Networks has tens of smaller customers who are starting to do AI pilots with the company that typically involve a few hundred GPUs; these customers go to Arista Networks for AI trials because they want best-of-breed reliability and performance

We have tens of smaller customers who are starting to do AI pilots…

…They’re about to build an AI cluster. It’s a reasonably small size, not classified in thousands or 10 thousands. But you’ve got to start somewhere. So they started about a few hundred GPUs, would you say?…

…The AI cloud we talked about, they tend to be smaller, but it’s a representation of the confidence the customer has. They may be using other GPUs, servers, et cetera. But when it comes to the mission critical networks, they’ve recognized the importance of best-of-breed reliability, availability, performance, no loss and the familiarity with the data center is naturally leading to pilots and trials on the AI side with us.

Arista Networks’ management classifies its TAM (total addressable market) within AI as how much of Infiniband will move to Ethernet and it’s far larger than the AI-related revenue of $750 million that management has guided for in 2025

The TAM is far greater than the $750 million we’ve signed up for. And remember, that’s early years. But that can consist of our data center TAM. Our AI TAM, which we count in a more narrow fashion as how much of InfiniBand will move to Ethernet on the back end. We don’t count the AI TAM that’s already in the front end, which is part and parcel of our data center.

Arista Networks’ management continues to see its large customers preferring to spend on AI, but is also seeing classic cloud continue to be an important part of its business and they believe the demand for classic cloud infrastructure will eventually rebound once the AI models are more established

We saw that last year. We saw that there was a lot of pivot going on from the classic cloud, as I like to call it, to the AI in terms of spend. And we continue to see favorable preferences to AI spend in many of our large cloud customers. Having said that, at the same time, simultaneously, we are going through a refresh cycle where many of these customers are moving from 100 to 200 or 200 to 400 gig. So while we think AI will grow faster than cloud, we’re betting on classic cloud continuing to be an important aspect of our contributions…

… I would say there’s such a heavy bias towards — in the Cloud Titans towards training and super training and the bigger and better the GPUs, the billion parameters, the OpenAI, ChatGPT and [indiscernible] that you’re absolutely right that at some level, the classic cloud, what you call traditional, I’m still calling classic, is a little bit neglected last year and this year. Having said that, I think once the training models are established, I believe this will come back, and it will sort of be a vicious cycle that feeds on each other. But at the moment, we’re seeing more activity on the AI and more moderate activity on the cloud.

Arista Networks’ management thinks that as AI networking moves towards Ethernet, it will be difficult to distinguish between front-end and back-end networks

It’s going to become difficult to distinguish the back end from the front end when they all move to Ethernet. For this AI center, as we call it, is going to be a conglomeration of both the front and the back. So if I were to fast forward 3, 4 years from now, I think the AI center is a supercenter of both the front end and the back end. So we’ll be able to track it as long as there’s GPUs and strictly training use cases. But if I were to fast forward, I think there may be many more edge use cases, many more inference use cases and many more small-scale training use cases which will make that distinction difficult to make.

Arista Networks’ management sees NVIDIA more as a friend than a competitor despite NVIDIA trying to compete with the company with the Spectrum-X switches; management rarely sees Spectrum-X as a competing technology in the deals Arista Networks is working on; management feels good about Arista Networks’ win rate

[Question] If you’re seeing Spectrum-X from NVIDIA? And if so, how you’re doing against it?

[Answer] When you say competitive environment, it’s complicated with NVIDIA because we really consider them a friend on the GPUs as well as the mix, so not quite a competitor. But absolutely, we will compete with them on the Spectrum switch. We have not seen the Spectrum except in one customer where it was bundled. But otherwise, we feel pretty good about our win rate and our success for a number of reasons, great software, portfolio of products and architecture that has proven performance, visibility features, management capabilities, high availability. And so I think it’s fair to say that if a customer were bundling with their GPUs, then we wouldn’t see it. If a customer were looking for best of breed, we absolutely see it and win it.

When designing GPU clusters for AI, a network design-centric approach has to be taken

If you look at an AI network design, you can look at it through 2 lenses, just through the compute, in which case you look at scale up and you look at it strictly through how many processes there are. But when we look at an AI network design, it’s a number of GPUs or XTUs per workload. Distribution and location of these GPUs are important. And whether the cluster has multiple tenants and how it’s divvied up between the host, the memory, the storage and the wide area plays a role, and the optimization to make on the applications for the collective communication libraries for specific workloads, levels of resilience, how much redundancy you want to put in, active, link base, load balancing, types of visibility. So the metrics are just getting more and more. There are many more commutations in combination. But it all starts with number of GPUs, performance and billions of parameters. Because the training models are definitely centered around job completion time. But then there’s multiple concentric circles of additional things we have to add to that network design. All this to say, a network design-centric approach has to be taken for these GPU clusters. Otherwise, you end up being very siloed

Arista Networks’ management is seeing huge clusters of GPUs – in the tens of thousands to hundreds of thousands – being deployed in 2025

Let me just remind you of how we are approaching 2024, including Q4, right? Last year, trials. So small, it was not material. This year, we’re definitely going into pilots. Some of the GPUs, and you’ve seen this in public blogs published by some of our customers have already gone from tens of thousands to 24,000 and are heading towards 50,000 GPUs. Next year, I think there will be many of them heading into tens of thousands aiming for 100,000 GPUs. So I see next year as more promising.

ASML (NASDAQ: ASML)

ASML’s management sees no change to the company’s outlook for 2024 from what was mentioned in the 2023 Q4 earnings call and 2024 Q1 earnings call, with AI-related applications still driving demand

Our outlook for the full year 2024 has not changed. We expect a revenue similar to last year. As indicated before, and based on our current guidance, the second half of the year is expected to be significantly higher than the first half. This is in line with the industry’s continued recovery from the downturn. Our guidance on market segments is similar to what we’ve stated in previous quarters…

……We currently see strong developments in AI driving most of the industry recovery and growth, ahead of other end market segments.

ASML’s management sees AI driving the majority of recovery in the semiconductor industry in both Logic and Memory chips; AI’s positive effects on semiconductor industry demand will start showing up in 2025 and management expects that to continue into 2026; Memory chips used in AI require high-bandwidth memory and so have higher density of DRAM; ASML’s management sees other non-AI segments as being behind in terms of recovery, but they do expect recovery eventually

We currently see strong developments in AI driving most of the industry recovery and growth, ahead of other end market segments…

… I think AI is driving, I would say, right now, the biggest part of the recovery. This is true for Logic. This is true for Memory. Roger just commented on Logic. I think you know that for high-bandwidth memory, those products drive more demand, more of a wafer demand because we are looking basically at a higher density of DRAM on those products. And we look at something that, of course, will take course over several months. So we started to see the positive effect of that for 2025. We expect that to continue into 2026, both for Memory and for Logic. And at some point of time, I also mentioned that maybe the other segments are a bit behind in terms of recovery.

So a lot of the capacity today, either Logic or DRAM capacity will be [indiscernible] those AI product. As the other segments recover, we also expect potentially some capacity to be needed there. 

ASML’s management thinks DRAM for AI memory chips will continue to see an increasing use of EUV lithography at each technology node; management also see opportunity for DRAM to use High-NA EUV lithography systems in 2025 or 2026

On DRAM, so I think there also, I think I’ll be very consistent with the information we have shared with you previously. So we see on there an increase of EUV use on every node. I think this is a trend that continue at least in the foreseeable future. Of course, it’s always more difficult to make forecast on nodes or technology that are still being defined by a customer. But that logic is still in place. I think you have seen also in DRAM that at this point of time, all customers are using EUV in production. I think the last customer was very public about that recently.

ASML’s management is not seeing much revenue made on AI at the moment, but it’s still seeing a lot of investment made for AI and these investments require a lot of semiconductor manufacturing capacity

I think what we have seen with AI is a major investment from many companies in supercomputer and the ability basically to train model. What we still miss in AI, I think, is the emergence of end product. So I think today, there’s not much revenue made on AI. There’s just a lot of investment. What we see is that still that investment require a lot of capacity. I think you have seen some of our customers announcing also more capacity to be built before 2028.

Coupang (NYSE: CPNG)

Coupang’s Product Commerce segment had sequential and year-on-year improvement in gross profit in 2024 Q2, driven partly by the use of AI technologies

Product Commerce gross profit increased 26% year-over-year to over $1.9 billion, and a record gross profit margin of 30.3%. This represents a 310 basis points improvement over last year and 200 basis points over last quarter. Our margin improvement this quarter was driven by strong growth rates in categories with higher margin composition, as well as higher efficiencies across operations, including benefits from greater utilization of automation and technology, including AI. We also continue to benefit from further optimization in our supply chain, and the scaling of margin accretive offerings.

Datadog (NASDAQ: DDOG)

Datadog’s management classifies digital natives as SMBs and mid-market companies, and within digital natives, the AI natives are inflecting in usage growth that others are not

I would add that the digital natives are largely SMB and mid-market, they’re not enterprise. And even when you look at the digital native, there’s two stories, depending on whether you talk about the AI natives or the others. The AI natives are inflecting in a way that the others are not at this point. So today, we see this higher growth from AI natives and from traditional enterprises. And stable growth, but not accelerating, from the rest of the pack.  

Datadog’s management has announced general availability of LLM Observability for generative AI for companies to monitor, troubleshoot, and secure LLM (large language model) applications; WHOOP and AppFolio are two early adopters of LLM Observability; it’s still very early days for the LLM Observability product; management thinks a good proxy for the future demand for LLM Observability is the growth of the model providers and the AI-native companies; management expects the LLM market to change a lot over time because it’s still nascent; in order of LLMs to work, they need to be connected to other applications and it’s at that point where management thinks the LLMs need observability; customers that are currently using LLM Observability also use Datadog for the rest of their technology stack and it does not make sense for the customers to operate their LLM applications in isolation

In the next-gen AI space, we announced the general availability of LLM Observability, which application developers and machine learning engineers to efficiently monitor, troubleshoot and secure LLM applications. With LLM Observability, companies can accelerate the deployment of AI applications into production environments and reliably operate and scale them…

… It’s still early. We do see customers that are going increasingly into production, and we have a few of those. I mean, we named a couple as early customers of LLM Observability. I think the two we named were WHOOP, the fitness band; and AppFolio. And we see many more that are lining up and then are going to do that. But in the grand scheme of things, looking at the whole market, it’s still very early. I would say the best proxy you can get from the future demand there is the growth of the model providers and the AI natives because they tend to be the ones that currently are being used to provide AI functionality into other applications and largely in production environment. And so I always said they are the harbinger of what’s to come…

… [Question] When people are thinking about bringing on LLMs into their organization, do they want the observability product in place already? Or are they testing out LLMs and then bringing you on after the fact?

[Answer] We expect this market to change a lot over time because it is far from being mature. And so a lot of the things that might happen today in a certain way might happen 2 years in a very, very different form. That being said, the way it works typically is customers build applications using developer tools, and there’s a whole industry that has emerged around developer tools for — and playgrounds and things like that for LLM. And so they use not one, but 100 different things to do that, which is fairly similar to what you might find on the IDE side or code editor side for the more traditional development, which is lots of different, very fragmented environment on that side. When they start connecting the LLM to the rest of the application, then they start to need like visibility that includes the other components because the LLM doesn’t work in a vacuum, it’s plugged into a front end. It works with authentication and security. It works with — connects to other system databases in other services to get the data. And at that point, they need it to be integrated with the rest of the observability. For the customers that use our LLM Observability product, they use us for the rest — all the rest of their stack. And it would make absolutely no sense for them to operate their LLM in isolation completely separately and not have the visibility across the whole applications. So it’s — at that point, it’s a no-brainer that they need everything to be integrated in production.    

Datadog’s management has expanded Bits AI, Datadog’s AI copilot, with new capabilities, such as the ability to perform autonomous investigations

We also expanded Bits AI with new capabilities. As a reminder, Bits AI is a Datadog built-in AI copilot. In addition to being able to summarize incidents and answer questions, we previewed at DASH, the ability for Bits AI to operate as an agent and perform autonomous investigations. With this capability, this AI proactively surfaces key information and performs complex tasks such as investigating alerts and coordinating — response.

Datadog’s management is hearing from all of Datadog’s customers that they are ramping experiments with AI with the goal of delivering business value with the technology; currently, 2,500 Datadog customers are using one or more of Datadog’s AI integrations for visibility into their use of AI; AI-native customers accounted for 4% of Datadog’s ARR in June 2024 (was 3.5% 2024 Q1); management thinks the percentage of ARR from AI-native customers will lose its relevance over time as AI usage becomes more widespread

Taking a step back and looking at our customer base, we continue to see a lot of excitement around AI technologies. All customers are telling us that they are leveling up on AI and ramping experimentations with the goal of delivering additional business value with AI. And we can see them doing this. Today, about 2,500 customers use one or more of our AI integrations to get visibility into their increasing use of AI. We also continue to grow our business with AI-native customers. which increased to over 4% of our ARR in June. We see this as a sign of the continued expansion of this ecosystem and of the value of using Datadog to monitor the product environment. I will note that over time, we think this metric will become less relevant as AI usage and production broadens beyond this group of customers.

Datadog’s management recently announced Toto, Datadog’s first foundational model for time-series forecasting; Toto delivered state-of-the-art performance on all 11 benchmarks; Toto’s capabilities come from the quality of Datadog’s training dataset; management sees Toto’s existence as evidence of the company’s ability to train, build, and incorporate AI models into its platform

We announced Toto, our first foundational model for time-series forecasting, which delivered state-of-the-art performance on all 11 benchmarks. In addition to the technical innovations devised by our research team, TOTO derives its record performance from the quality of our training dataset and points to our unique ability to train, build and incorporate AI models into a platform that will meaningfully improve operations for our customers.

Datadog’s management continues to believe that digital transformation, cloud migration, and AI adoption are long-term growth drivers of Datadog’s business

Overall, we continue to see no change to the multiyear trend towards digital transformation and cloud migration. We are seeing continued experimentation with new technologies, including next-gen AI, and we believe this is just one of the many factors that will drive greater use of the cloud and next-gen infrastructure.

Datadog’s management thinks the emergence of AI has led to large enterprises realising they need to be on the cloud sooner rather later; management sees a lot of growth in the cloud migration of enterprises as it’s really early in their transition

Some of the strengths we see today has to do with the fact that, to serve their — in part to — the emergence of AI has reaffirmed for them the need to go to the cloud sooner rather than later. So they can build the right kind of applications, they have the right kind of data available to give those applications…

…I’d point you to the numbers we shared, I think, 2 quarters ago in terms of our enterprise penetration and the average size of our contracts with enterprises, which are still fairly small. Like there’s a lot of runway there. And the growth of those accounts is not predicated on the growth of the enterprise themselves. They’re still early in their transformation.

Fiverr (NYSE: FVRR)

Fiverr’s management is deepening the integration of Neo, the company’s AI assistant, into its marketplace experience; management realised that not everyone wants the outright chatbot experience on its marketplace, so Neo only pops up when friction arises to provide guidance for buyers who are navigating Fiverr’s catalogue of talent; management wants Neo to be a personal assistant throughout the Fiverr purchasing experience and also answer buyers’ questions

The second theme of our Summer Product Release is deepening the integration of Neo, Fiverr’s AI tool throughout the market-based experience. As Gen-AI applications quickly shift consumers’ Internet behavior and expectations, we want to stay ahead of the curve to build a more personable experience on Fiverr. At the same time, tests and data in the past 6 months have shown that not everyone prepares the outright chatbot experience when it comes to shopping. So, our strategy for Neo is to incorporate it as an assistance throughout the funnel to help customers when friction arises. For search, Neo provides the guidance you need to navigate Fiverr’s massive catalog of services and talent. And it is trained to understand customers’ past transactions and preference to provide the most relevant recommendations. When it comes to project briefing, having Neo is like having a strategist by your side. It transforms customers’ ideas into a structured brief document that not only looks good, but also delivers better business results. Neo can also help customers write more detailed reviews faster by generating content based on transactions and providing language assistance…

…The experimentation that we’ve done with Neo as a personal assistant within the inbox, which is the — which was the first version of doing it, taught us a lot about how our customers are actually using it and how it improves the conversion in briefing. It allows buyers to complete, and it leads to higher conversion as a result. And so, the idea here is that we’re graduating Neo to get out of the inbox and essentially being integrated in all of our experience. Right now, it’s being rolled out gradually because we want to test its accuracy and performance. But essentially, you can fund it as a personal assistant throughout the experience. So, it allows customers to search better, to be more accurate about their needs, and as a result get much higher quality match.

But it also has awareness about where it exists. So, if you’re looking at a specific page, you can ask questions about that page. So, it helps people make decisions and get to what they’re looking for better. The same goes with the integration in briefing. If customers have a brief premade then they can just upload it, and we help make that brief even better. But if they don’t, then the technology that is behind Neo actually helps them write a better, more accurate brief and again, as a result of that, get matched with a much more specific cohort of potential talent that can do the job.

Fiverr’s management continues to believe that AI will be a multiyear tailwind for the company and that AI will have a net positive impact on the company’s business; the deterioration seen in the simple services categories has improved, for whatever reasons (unsure if it’s a one-off event from low base, as management also spoke about the low-base effect); around 20% of Fiverr’s GMV comes from simple jobs 

We are in the early innings of unleashing the full potential of AI in our marketplace, and we believe it will be a multiyear tailwind for us to drive product innovation and growth…

…We also see AI continuing to have a net positive impact on our business. It is important to note that we are starting to see stabilizing and improving trends in simple services…

…Now several quarters in, we are actually seeing that in our — we’re seeing this in our data. So, for example, writing and translation as a vertical is the vertical with the biggest exposure to AI impact. In Q2, we’re actually seeing traffic in that vertical improved 10 percentage points in terms of year-over-year growth rate compared to Q1…

…That said, with us now opening professions catalog and hourly contracts this will open up new funnels and create growth opportunities, especially for complex services categories. And remember that we have over 700 categories. So, our exposure to specific categories is relatively low and seasonal trends in category spend are a regular thing in our line of business…

…When we think about the overall mix complex is in the mid-30s of GMV and simple is about 20%.

Mastercard (NYSE: MA)

Mastercard’s management intends to further embed AI into Mastercard’s value-added services, particularly in data analytics, fraud, and cybersecurity, because they are seeing companies asking for these solutions; the embedding of AI into the value-added services portfolio does not involve changing the existing portfolio, but augmenting them with a higher weightage to AI

We will also enhance and expand our value-added services, such as in data analytics, fraud and cybersecurity particularly as we further embed AI into our products and services…

…It’s pretty clear that on the services side, as far as the areas of focus are concerned, we continue to be guided by underlying strong secular trends, and one of that is for really any of our corporate partners and B2B partners that they want to make sense of their enterprise data and make better decisions. And how do we do that? We do that by leveraging our artificial intelligence solutions, our set of assistants, a set of fine-tuning, how they could have more personalized suggestions to their end consumers, et cetera, et cetera. That’s one part, help our customers make better decisions, not changing, but very specific solutions with a higher weightage to AI.

And then on the security side and the cybersecurity side, all of this data has to be kept safe. We kept saying that for years. That’s a strong secular trend in itself and making sure that we fine-tune our solutions here. We’ve got to move faster because the bad guys are also moving faster, and they have the similar technology tools at their hand now. So leveraging artificial intelligence, an example I gave last quarter around Decision Intelligence Pro, that’s predicting what is the next card that might be frauded, before it actually happens. Those kind of solutions provide significant lift to our customers in terms of preventing fraud, obviously giving peace of mind to their consumers and overall helping our business, and it’s a close link to our payments — underlying payments business.

Mastercard has been using AI technology successfully for the better part of a decade, in areas such as fraud prevention; management thinks generative AI gives the opportunity for Mastercard to understand more data faster; management has used generative AI to create artificial data sets to train Mastercard’s discriminative AI models; management has also used generative AI to build a new product, such as Decision Intelligence Pro; Decision Intelligence Pro brings a 20% improvement in fraud prediction; management believes that generative AI will increase in penetration within Mastercard’s fraud and cybersecurity products 

 AI isn’t actually anything new for us. So we’ve — for the better part of a decade, we’ve been using AI. This is a discrete machine learning technology to really predict where is the next problem, and analyze data of — that we have and the data that our customers have to prevent fraud. So that’s been very successful.

As far as generative AI is concerned, evolving technology here, there’s obviously an opportunity for us to understand more data in a quicker way. And we have used that initially to train our AI models, our discriminative AI models using generative AI to create artificial data set. So that was the first step. And then we went into putting out a new set of products. I mentioned Decision Intelligence Pro. Decision Intelligence is a product that we’ve had for a long time, machine learning driven that was predicting fraud outcomes and now we’re using more data sets to — that are externally available, stolen card data and so forth, to understand where fraud vulnerabilities might be. The lift is tremendous, 20%, we see in terms of effectiveness out of that product. So we start to see demand for the whole reason on the vulnerabilities that I talked about…

…I believe that the penetration of generative AI and our fraud and cybersecurity product set will only expand. 

Mercado Libre (NASDAQ: MELI)

MercadoLibre has been putting a lot of resources into AI and generative AI; management sees many ways AI can help the commerce business, such as producing better ways for consumers to look at product reviews, enhance product pictures, generate seller-responses when sellers are unable to, and improve the product search experience for consumers; MercadoLibre has 16,000 developers and they are using AI to improve productivity; MercadoLibre is using AI inc customer support to respond more cost-effectively and more accurately

We have been — put a lot of resources into AI and GenAI throughout the company, really. We don’t have a centralized department of AI, but all of our different business units…

… On the commerce side, obviously, we are using AI to help us with recommendations, as you mentioned, but more important than that on reviews, for instance, that in the past, you have to — if you were to review a product, you have to go through many different views, now we can consolidate that into a more efficient way of communicating the qualities, the prospects of a particular product pictures, as you know, our pictures that publish might not be the quality that we are expecting from our merchants, and we can improve those with answers from sellers is another good example in the past, if you were to buy something at 2 AM in the morning, you’ll have to wait until the next day to get an answer that obviously affected significantly the conversion of the product. Now we can respond right away with using GenAI models…

…On the developer side, we have 16,000 developers, which are also using AI tools to improve productivity and that also generating some improvements and efficiencies in the way we deploy products throughout the company. And I think 1 of the most important projects that we have is on CX, customer experience and customer support by which we are also applying AI tools that will help us to not only respond more efficiently in terms of cost, but also be more accurate in terms of the way we manage those issues. These are some examples, but there are many others…

… You asked about search and where you’re seeing technical and bedding to power search that technical — turn search into something more semantic. So it’s easier to try to send the users to what they’re looking for.

Meta Platforms (NASDAQ: META)

Meta’s AI work continues to improve quality of recommendations on Facebook and Instagram, and drives engagement; the more general recommendation models Meta develops, the better the content recommendations get; Meta rolled out a unified video recommendation service across Facebook in 2024 Q2 for Reels, longer videos, and Live; Meta’s unified AI systems had already increased engagement on Facebook Reels more than Meta’s shift from using CPUs to GPUs; management wants to eventually have a single, unified AI recommendation system for all kinds of content across Meta’s social apps; the unified video recommendation service has encouraging early results, and management expects the relevance of video recommendations to increase

Across Facebook and Instagram, advances in AI continue to improve the quality of recommendations and drive engagement. And we keep finding that as we develop more general recommendation models, content recommendations get better. This quarter we rolled out our full-screen video player and unified video recommendation service across Facebook — bringing Reels, longer videos, and Live into a single experience. This has allowed us to extend our unified AI systems, which had already increased engagement on Facebook Reels more than our initial move from CPUs to GPUs did. Over time, I’d like to see us move towards a single, unified recommendation system that powers all of the content including things like People You May Know across all of our surfaces. We’re not there, so there’s still upside — and we’re making good progress here…

…On Facebook, we are seeing encouraging early results from the global roll-out of our unified video player and ranking systems in June. This initiative allows us to bring all video types on Facebook into one viewing experience, which we expect will unlock additional growth opportunities for short-form video as we increasingly mix shorter videos into the overall base of Facebook video engagement. We expect the relevance of video recommendations will continue to increase as we benefit from unifying video ranking across Facebook and integrating our next generation recommendation systems. These have already shown promising gains since we began using the new systems to support Facebook Reels recommendations last year. We expect to expand these new systems to support more surfaces beyond Facebook video over the course of this year and next year

In the past, advertisers would tell Meta the specific audience they wanted to reach, but over time, Meta could predict the interested-audience better than the advertisers could, even though the advertisers still needed to come up with collateral; management thinks that AI will generate personalised collateral for advertisers in the coming years and all the advertiser needs to do is to tell Meta a business objective and a budget, and Meta will handle everything else; Meta’s first generative AI ad features, such as image expansion and text generation, were used by more than 1 million advertisers in June 2024; Meta rolled out full image generation capabilities in Advantage+ in May 2024

It used to be that advertisers came to us with a specific audience they wanted to reach — like a certain age group, geography, or interests. Eventually we got to the point where our ads system could better predict who would be interested than the advertisers could themselves. But today advertisers still need to develop creative themselves. In the coming years, AI will be able to generate creative for advertisers as well — and will also be able to personalize it as people see it. Over the long term, advertisers will basically just be able to tell us a business objective and a budget, and we’re going to go do the rest for them. We’re going to get there incrementally over time, but I think this is going to be a very big deal…

…We’ve seen promising early results since introducing our first generative AI ad features – image expansion, background generation, and text generation – with more than one million advertisers using at least one of these solutions in the past month. In May, we began rolling out full image generation capabilities into Advantage+ creative, and we’re already seeing improved performance from advertisers using the tool. 

Meta’s management thinks that Meta AI, the company’s AI assistant feature, will be the most used AI assistant by end-2024; Meta AI is improving in intelligence and features quickly, and seems on track to be an important service; Meta AI’s current use cases include searching for information, role-playing difficult conversations, and creating images, but new use cases are likely to emerge; Meta AI has been used for billions of queries thus far; Meta AI has helped with WhatsApp retention and engagement; India has become the largest market for Meta AI; Meta AI is now available in 20 countries and 8 languages; management thinks that people who bet on the early indicators of Meta tend to do pretty well, and Meta AI is one of those early indicators that are signalling well; management wants to build a lot more functionality into Meta AI, but that will take a few years

Last quarter we started broadly rolling out our assistant, Meta AI, and it is on track to achieve our goal of becoming the most used AI assistant by the end of the year. We have an exciting roadmap ahead of things that we want to add, but the bottom line here is that Meta AI feels on track to be an important service and it’s improving quickly both in intelligence and features. Some of the use cases are utilitarian, like searching for information or role-playing difficult conversations before you have them with another person, and other uses are more creative, like the new Imagine Yourself feature that lets you create images of yourself doing whatever you want in whatever style you want. And part of the beauty of AI is that it’s general, so we’re still uncovering the wide range of use cases that it’s valuable for…

…People have used Meta AI for billions of queries since we first introduced it. We’re seeing particularly promising signs on WhatsApp in terms of retention and engagement, which has coincided with India becoming our largest market for Meta AI usage. You can now use Meta AI in over 20 countries and eight languages, and in the US we’re rolling out new features like Imagine edit, which allows people to edit images they generate with Meta AI…

… I think that the people who bet on those early indicators tend to do pretty well, which is why I wanted to share in my comments the early indicator that we had on Meta AI, which is, I mean, look, it’s early…

…I was talking before about we have the initial usage trends around Meta AI but there’s a lot more that we want to add, things like commerce and you can just go vertical by vertical and build out specific functionality to make it useful in all these different areas are eventually, I think, what we’re going to need to do to make this just as — to fulfill the potential around just being the ideal AI assistant for people. So it’s a long road map. I don’t think that this stuff is going to get finished in the next couple of quarters or anything like that. But this is part of what’s going to happen over the next few years as we build something that will, I think, just be a very widely used service. So I’m quite excited about that.

Meta’s management recently launched AI Studio, which allows anyone to create AIs that people can interact with; AI Studio is useful for creators who want to engage more with their communities, but can also be useful for anyone who wants to build their own AI agents, including businesses; management thinks every business in the future will have its own AI agent for customer interactions that drives sales and reduces costs; management expects Business AI agents to dramatically accelerate Meta’s business messaging revenue when the feature reaches scale

This week we launched AI Studio, which lets anyone create AIs to interact with across our apps. I think that creators are especially going to find this quite valuable. There are millions of creators across our apps — and these are people who want to engage more with their communities and their communities want to engage more with them — but there are only so many hours in the day. So now they’re going to be able to use AI Studio to create AI agents that can channel them to chat with their community, answer people’s questions, create content, and more. So I’m quite excited about this. But this goes beyond creators too. Anyone is going to be able to build their own AIs based on their interests or different topics that they are going to be able to engage with or share with their friends.

Business AIs are the other big piece here. We’re still in alpha testing with more and more businesses. The feedback we’re getting is positive so far. Over time I think that just like every business has a website, social media presence, and an email address, in the future I think that every business is also going to have an AI agent that their customers can interact with. Our goal is to make it easy for every small business, and eventually every business, to pull all their content and catalog into an AI agent that drives sales and saves them money. When this is working at scale, I expect it to dramatically accelerate our business messaging revenue.

The Llama family of foundation models is the engine that powers all of Meta’s AI-related work; in 2024 Q2, Meta released Llama 3.1, the first frontier-level open source model, and other new and industry-leading small and medium models; the Llama 3.1 405B model has better cost performance compared to leading closed models; management thinks Llama 3.1 will mark an inflection point for open source AI becoming the industry standard; Meta is already working on Llama 4 and management is aiming for it to be the most advanced foundation AI model when released in 2025; the Llama models are well-supported by the entire cloud computing ecosystem

The engine that powers all these new experiences is the Llama family of foundation models. This quarter we released Llama 3.1, which includes the first frontier-level open source model, as well as new and industry-leading small and medium-sized models. The 405B model has better cost performance relative to the leading closed models, and because it’s open, it is immediately the best choice for fine-tuning and distilling your own custom models of whatever size you need. I think we’re going to look back at Llama 3.1 as an inflection point in the industry where open source AI started to become the industry standard, just like Linux is…

…We’re already starting to work on Llama 4, which we’re aiming to be the most advanced in the industry next year…

… Part of what we’re doing is working closely with AWS, I think, especially did great work for this release. Other companies like Databricks, NVIDIA, of course, other big players like Microsoft with Azure, and Google Cloud, they’re all supporting this. And we want developers to be able to get it anywhere. I think that’s one of the advantages of an open source model like Llama is — it’s not like you’re locked into 1 cloud that offers that model, whether it’s Microsoft with OpenAI or Google with Gemini or whatever it is, you can take this and use it everywhere and we want to encourage that. So I’m quite excited about that.

Meta’s management is planning for the AI compute needs of the company for the next several years; management thinks the compute requirements for training Llama 4 will likely be 10x that of Llama 3, and future models will require even more; given long lead times to build compute capacity, management would rather risk overbuilding than being too late in realising there’s a shortfall; even as Meta builds compute capacity, management still remains focused on cost efficiency

We’re planning for the compute clusters and data we’ll need for the next several years. The amount of compute needed to train Llama 4 will likely be almost 10x more than what we used to train Llama 3 — and future models will continue to grow beyond that. It’s hard to predict how this will trend multiple generations out into the future, but at this point I’d rather risk building capacity before it is needed, rather than too late, given the long lead times for spinning up new infra projects. And as we scale these investments, we’re of course going to remain committed to operational efficiency across the company…

A few years ago, management thought holographic AR (augmented reality) technology would be ready before smart AI, but the reverse has happened; regardless, Meta is still well positioned for this reverse order; because of AI, Meta’s smart glasses continue to be a bigger hit than management expected and supply cannot keep up with demand; Meta will continue to partner EssilorLuxottica for the long term to build its smart glasses

A few years ago I would have predicted that holographic AR would be possible before smart AI, but now it looks like those technologies will actually be ready in the opposite order. We’re well-positioned for that because of the Reality Labs investments that we’ve already made. Ray-Ban Meta glasses continue to be a bigger hit sooner than we expected — thanks in part to AI. Demand is still outpacing our ability to build them, but I’m hopeful we’ll be able to meet demand soon. EssilorLuxottica has been a great partner to work with on this, and we’re excited to team up with them to build future generations of AI glasses as we continue to build our long term partnership.

AI is playing an increasingly important role in improving Meta’s marketing performance; the AI-powered Meta Lattice ad ranking architecture continued to drive ad performance and efficiency gains in 2024 Q2; Advantage+ Shopping campaigns are driving 22% higher return on ad spend for US advertisers; advertiser adoption of Meta’s advertising automation tools continue to expand; Meta has continued to increase the capabilities of Advantage+, such as expanding conversion types, and helping advertisers automatically select which ad format to serve after they upload multiple images and videos; Meta rolled out full image generation capabilities in Advantage+ in May 2024

The second part of improving monetization efficiency is enhancing marketing performance. We continue to be pleased with progress here, with AI playing an increasingly central role. We’re improving ad delivery by adopting more sophisticated modeling techniques made possible by AI advancements, including our Meta Lattice ad ranking architecture, which continued to provide ad performance and efficiency gains in the second quarter. We’re also making it easier for advertisers to maximize ad performance and automate more of their campaign set up with our Advantage+ suite of solutions. We’re seeing these tools continue to unlock performance gains, with a study conducted this year demonstrating 22% higher return on ad spend for US advertisers after they adopted Advantage+ Shopping campaigns. Advertiser adoption of these tools continues to expand, and we’re adding new capabilities to make them even more useful. For example, this quarter we introduced Flexible Format to Advantage+ Shopping, which allows advertisers to upload multiple images and videos in a single ad that we can select from and automatically determine which format to serve, in order to yield the best performance. We have also now expanded the list of conversions that businesses can optimize for using Advantage+ Shopping to include an additional 10 conversion types, including objectives like “add to cart”…

…In May, we began rolling out full image generation capabilities into Advantage+ creative, and we’re already seeing improved performance from advertisers using the tool. 

Monetisation for Meta’s AI products such as Meta AI or AI Studio will take years because management is following the same playbook they have had for years, which is to start a product, then take time to scale the product to a billion users before monetising; Meta’s management is a little different from other companies in terms of how they think about the time needed to monetise products

We have a relatively long business cycle of starting a new product, scaling it to something that reaches 1 billion people or more and only then really focusing on monetizing at scale. So realistically, for things like Meta AI or AI Studio, I mean, these are things that I think will increase engagement in our products and have other benefits that will improve the business and engagement in the near term. But before we’re really talking about monetization of any of those things by themselves, I mean, I don’t think that anyone should be surprised that I would expect that, that will be years, right?…

…And I think that, that’s something that is a little bit different about Meta in the way we build consumer products and the business around them than a lot of other companies that ship something and start selling it and making revenue from it immediately. So I think that’s something that our investors and folks thinking about analyzing the business, if needed, to always grapple with is all these new products, we ship them and then there’s a multiyear time horizon between scaling them and then scaling them into not just consumer experiences but very large businesses.

Meta’s ongoing capex investments in AI infrastructure is informed by the strong returns management has seen and expect to achieve in the future; management expects the returns from generative AI to take some time to appear, but they see signification monetisation opportunities that could be unlocked through the AI investments; Meta’s capital expenditures for AI infrastructure are done with flexibility in mind so that AI training capacity can also be redirected to generative AI inference and its ranking and recommendation systems, if needed; management is focused on improving cost efficiency of its AI workloads over time; Meta’s AI capex come in 2 buckets, core AI and generative AI (genAI), which are built to be fungible if needed; the core AI bucket is much more mature in driving revenue for Meta and management takes an ROI (return on investment) approach; the gen AI bucket is much earlier in revenue-generation-maturity but is expected to open up new revenue opportunities over time to deliver that ROI; it’s difficult for management to plan for Meta’s long-term capex trajectory

Our ongoing investment in core AI capacity is informed by the strong returns we’ve seen, and expect to deliver in the future, as we advance the relevance of recommended content and ads on our platforms. While we expect the returns from generative AI to come in over a longer period of time, we are mapping these investments against the significant monetization opportunities that we expect to be unlocked across customized ad creative, business messaging, a leading AI assistant, and organic content generation. As we scale generative AI training capacity to advance our foundation models, we will continue to build our infrastructure in a way that provides us with flexibility in how we use it over time. This will allow us to direct training capacity to gen AI inference, or to our core ranking and recommendation work when we expect that doing so would be more valuable. We will also continue our focus on improving the cost efficiency of our workloads over time…

… I would broadly characterize our AI investments into 2 buckets: core AI and gen AI. And the 2 are really at different stages as it relates to driving revenue for our businesses and our ability to measure returns. On our core AI work, we continue to take a very ROI-based approach to our investment here. We’re still seeing strong returns as improvements to both engagement and ad performance have translated into revenue gains, and it makes sense for us to continue investing here. Gen AI is where we’re much earlier, as Mark just mentioned in his comments. We don’t expect our gen AI products to be a meaningful driver of revenue in ’24. But we do expect that they’re going to open up new revenue opportunities over time that will enable us to generate a solid return off of our investment while we’re also open sourcing subsequent generations of Llama. And we’ve talked about the 4 primary areas that we’re focused here on the gen AI opportunities to enhance the core ads business, to help us grow in business messaging, the opportunities around Meta AI, and the opportunities to grow core engagement over time.

The other thing I would say is, we’re continuing to build our AI infrastructure with fungibility in mind so that we can flex capacity where we think it will be put to best use. The infrastructure that we build for gen AI training can also be used for gen AI inference. We can also use it for ranking and recommendations by making certain modifications like adding general compute and storage. And we’re also employing a strategy of staging our data center sites at various phases of development, which allows us to flex up to meet more demand and less lead time if needed while limiting how much spend we’re committing to in the outer years…

…We haven’t really shared an outlook sort of on the longer-term CapEx trajectory. In part, infrastructure is an extraordinarily dynamic planning area for us right now. We’re continuing to work through what the scope of the gen AI road maps will look like over that time. Our expectation, obviously again, is that we are going to significantly increase our investments in AI infrastructure next year, and we’ll give further guidance as appropriate. But we are building all of that CapEx, again with the factors in mind that I talked about previously, thinking about both how to build it flexibly so we can deploy to core AI and gen AI use cases as needed…

… There’s sort of a whole host of use cases for the life of any individual data center ranging from gen AI training at its outset to potentially supporting gen AI inference to being used for core ads and content ranking and recommendation and also thinking through the implications, too, of what kinds of servers we might use to support those different types of use cases.

Microsoft (NASDAQ: MSFT)

Microsoft’s management sees the AI platform shift as involving both knowledge and capital-intensive investments, similar to the Cloud platform shift; as Microsoft goes through the AI platform shift, management is focused on product innovation, and using customer demand signals and time to value to manage the cost structure dynamically

 I want to offer some broader perspective on the AI platform shift. Similar to the Cloud, this transition involves both knowledge and capital-intensive investments. And as we go through this shift, we are focused on 2 fundamental things. First, driving innovation across a product portfolio that spans infrastructure and applications, so as to ensure that we are maximizing our opportunity while in parallel, continuing to scale our cloud business and prioritizing fundamentals, starting with security. Second, using customer demand signal and time to value to manage our cost structure dynamically and generate durable long-term operating leverage.

Azure’s share gains accelerated in FY2024 (fiscal year ended 30 June 2024), driven by AI; Azure grew revenue by 29% in 2024 Q2 (was 31% in 2024 Q1), with 8 points of growth from AI services (was 7 points in 2024 Q1); Azure’s AI business has higher demand than available capacity; 50% of Azure AI users are also using a data meter within Azure, which is excellent for Azure

Starting with Azure. Our share gains accelerated this year driven by AI…

…Azure and other cloud services revenue grew 29% and 30% in constant currency, in line with expectations and consistent with Q3 when adjusting for leap year. Azure growth included 8 points from AI services, where demand remained higher than our available capacity…

…AI doesn’t sit on its own, right? So it’s just for — we have a concept of design wins in Azure. So in fact, 50% of the folks who are using Azure AI are also using a data meter. That’s very exciting to us because the most important thing in Azure is to win workloads in the enterprise. And that is starting to happen. And these are generational things once they get going with you. So that’s, I think, how we think about it at least when I look at what’s happening on our demand side. 

Azure added new AI accelerators from both AMD and NVIDIA, and its own in-house Azure Maia chips; Azure also introduced its own Cobalt 100 CPUs

We added new AI accelerators from AMD and NVIDIA as well as our own first-party silicon Azure Maia and we introduced new Cobalt 100, which provides best-in-class performance for customers like Elastic, MongoDB, Siemens, Snowflake and Teradata.

Azure AI offers the most diverse selection of models for customers; Azure AI now has 60,000 customers and average spend per customer continues to grow; Azure OpenAI started to provide access to GPT-4o and GPT-4o Mini in 2024 Q2; Azure OpenAI is being used by companies from diverse industries; Phi-3 within Azure AI offers small language models that are already being used by a wide range of companies; Models as a Service within Azure AI offers access to third-party models including open-sourced models and it is being used by a diverse range of large companies; paid Models as a Service customers doubled sequentially

With Azure AI, we are building out the app server for the AI wave providing access to the most diverse selection of models to meet customers’ unique cost, latency and design considerations. All up, we now have over 60,000 Azure AI customers up nearly 60% year-over-year and average spend per customer continues to grow.  Azure OpenAI service provides access to best-in-class frontier models, including as of this quarter GPT-4o and GPT-4o mini. It’s being used by leading companies in every industry, including H&R Block, Suzuki, Swiss Re, Telstra as well as digital natives like Freshworks, Meesho and Zomato. With Phi-3, we offer a family of powerful small language models, which are being used by companies like BlackRock, Emirates, Epic, ITC, Navy Federal Credit Union and others. And with Models as a Service, we provide API access to third-party models, including as of last week, the latest from Cohere, Meta and Mistral. The number of paid Models as a Service customers more than doubled quarter-over-quarter, and we are seeing increased usage by leaders in every industry from Adobe and Bridgestone to Novo Nordisk and Palantir.

Microsoft Fabric, an AI-powered data platform, now has more than 14,000 customers (was more than 11,000 in 2024 Q1)

Microsoft Fabric, our AI-powered next-generation data platform, now has over 14,000 paid customers, including leaders in every industry from Accenture and Kroger to Rockwell Automation and Zeiss, up 20% quarter-over-quarter. And this quarter, we introduced new first of their kind, real-time intelligence capabilities in Fabric, so customers can unlock insights on high-volume, time-sensitive data.

GitHub Copilot is the most widely adopted AI-powered developer tool; 77,000 organisations have adopted GitHub Copilot in just over 2 years since its general availability and the number of organisations is up 180% from a year ago; GitHub Copilot is driving GitHub’s overall growth; GitHub’s annual revenue run rate is $2 billion and Copilot accounted for more than 40% of GitHub’s revenue growth in FY2024; GitHub Copilot alone is already a larger business than the entire GitHub when Microsoft acquired it in 2018

GitHub Copilot is by far the most widely adopted AI power developer tool. Just over 2 years since its general availability, more than 77,000 organizations from BBVA, FedEx and H&M to Infosys and Paytm have adopted Copilot up 180% year-over-year…

…Copilot is driving GitHub growth all up. GitHub annual revenue run rate is now $2 billion. Copilot accounted for over 40% of GitHub revenue growth this year and is already a larger business than all of GitHub was when we acquired it.

More than 480,000 organisations have used AI-features within Microsoft’s Power Platform (was more than 330,000 in 2024 Q1), and Power Platform has 48 million monthly active users (was 25 million in 2024 Q1)

We are also integrating generative AI across Power Platform, enabling anyone to use natural language to create apps, automate workflows or build a website. To date, over 480,000 organizations have used AI-powered capabilities in Power Platform, up 45% quarter-over-quarter. In total, we now have 48 million monthly active users of Power Platform, up 40% year-over-year.

The number of Copilot for Microsoft 365 users doubled sequentially; Copilot for Microsoft 365 customers increased 60% sequentially; number of customers for Copilot for Microsoft 365 with more than 10,000 seats doubled sequentially; Copilot Studio customers can build custom Copilots for agentic work; 50,000 organisations have used Copilot Studio

Copilot for Microsoft 365 is becoming a daily habit for knowledge workers as it transforms work, workflow and work artifacts. The number of people who use Copilot daily at work nearly doubled quarter-over-quarter as they use it to complete tasks faster, hold more effective meetings and automate business workflows and processes. Copilot customers increased more than 60% quarter-over-quarter. Feedback has been positive with majority of enterprise customers coming back to purchase more seats, all up the number of customers with more than 10,000 seats more than doubled quarter-over-quarter, including Capital Group, Disney, Dow, Kyndryl, Novartis, and EY alone will deploy Copilot to 150,000 of its employees and we are going further adding agent capabilities to Copilot. New Team Copilot can facilitate meetings and create an assigned task. And with Copilot Studio customers can extend Copilot for Microsoft 365 and build custom Copilots that proactively respond to data and events using their own first and third-party business data. To date, 50,000 organizations from Carnival Corporation, Cognizant and Eaton to KPMG, Majesco and McKinsey have used Copilot Studio, up over 70% quarter-over-quarter.

DAX Copilot has been purchased by more than 400 healthcare organisations to-date, up 40% sequentially; the number of AI-generated clinical reports have tripled

With DAX Copilot, more than 400 health care organizations, including Community Health Network, Intermountain, Northwestern Memorial Healthcare and Ohio State University Wexner Medical Center have purchased DAX Copilot to date, up 40% quarter-over-quarter and the number of AI-generated clinical reports more than tripled.

Microsoft introduced a new category of Copilot+ PCs in 2024 Q2; the Copilot+ PCs have a new system architecture design to deliver breakthrough AI experiences; early reviews are promising

When it comes to devices, we introduced our new category of Copilot+ PCs this quarter. They are the fastest, most intelligent Windows PCs ever. They include a new system architecture designed to deliver best-in-class performance and breakthrough AI experiences. We are delighted by early reviews, and we are looking forward to the introduction of more Copilot+ PCs powered by all of our silicon and OEM partners in the coming months.

More than 1,000 paid customers used Copilot for security ; Microsoft now has 1.2 million security customers and over 800,000 of them use 4 or more workloads, up 25% from a year ago

Over 1,000 paid customers used Copilot for security, including Alaska Airlines, Oregon State University, Petrofac, Wipro, WTW, and we are also securing customers’ AI deployments with updates to Defender and Purview. All up, we now have 1.2 million security customers over 800,000, including Dell Technologies, Deutsche Telekom, TomTom use 4 or more workloads, up 25% year-over-year. 

Combined revenue of Bing, Edge, and Copilot was up 19% year-on-year and management said Bing and Edge took share; management is applying generative AI to Bing to test a new generative search experience, whose aim is to create dynamic responses while still driving clicks to publishers

We are ensuring that Bing, Edge and Copilot collectively are driving more engagement and value to end users, publishers and advertisers. Our overall revenue ex-TAC increased 19% year-over-year and we again took share across Bing and Edge. We continue to apply Generative AI to pioneer new approaches to how people search and browse. Just last week, we announced we are testing a new generative search experience, which creates a dynamic response to users’ query while maintaining click share to publishers. 

Copilot for the web has created more than 12 billion images and did more than 13 billion chats to-date, up 150% since the start of 2024

We continue to drive record engagement with Copilot for the web, consumers have used Copilot to create over 12 billion images and conduct 13 billion chats to date, up 150% since the start of the calendar year.

Microsoft is using AI in its Performance Max advertising tool to create and optimise ads for advertisers, increasing their advertising ROI (return on investment)

We are helping advertisers increase their ROI, too. We have seen positive response to Performance Max, which uses AI to dynamically create and optimize ads and Copilot and Microsoft ad platform helps marketers create campaigns and troubleshoot using natural language.

Microsoft’s capex in 2024 Q2 (FY2024 Q4) and the whole of FY2024 are basically for AI and cloud, and it can be split roughly 50-50 into (1) data centers and (2) servers consisting of GPU/CPUs; management sees the capex for the data centers as providing support for monetisation over the next 15-plus years; the capex for GPUs and CPUs are driven by demand signals; the demand signals that management is seeing include Microsoft 365 Copilot demand, GitHub Copilot demand, and Azure AI growth; Microsoft can be spending on the data centres first, because they have long lead times, without spending on the GPUs and CPUs if the demand signals no longer persist, moreover, revenue growth will not be affected by the throttling of GPU/CPU spending; part of the capex is for AI training, but management will be scaling training only if they see demand; the capex on the data centres itself is really flexible because Microsoft has built a consistent architecture for its technological infrastructure

Capital expenditures, including finance leases, were $19 billion, in line with expectations and cash paid for PP&E was $13.9 billion. Cloud and AI-related spend represents nearly all of our total capital expenditures. Within that, roughly half is for infrastructure needs where we continue to build and lease data centers that will support monetization over the next 15 years and beyond. The remaining Cloud and AI-related spend is primarily for servers, both CPUs and GPUs to serve customers based on demand signals. For the full fiscal year, the mix of our Cloud and AI-related spend was similar to Q4…

…So when I think about what’s happening with M365 Copilot as perhaps the best Office 365 or M365 suite we have had, the fact that we’re getting recurring customers, so our customers coming back buying more seats. So GitHub Copilot now being bigger than even GitHub when we bought it. What’s happening in the contact center with Dynamics. So I would say — and obviously, the Azure AI growth, that’s the first place we look at. That then drives bulk of the CapEx spend, basically, that’s the demand signal because you got to remember, even in the capital spend, there is land and there is data center build, but 60-plus percent is the kit, that only will be bought for inferencing and everything else if there is demand signal, right? So that’s, I think, the key way to think about capital cycle even. The asset, as Amy said, is a long-term asset, which is land and the data center, which, by the way, we don’t even construct things fully, we can even have things which are semi-constructive, we call Kohl’s shelves and so on. So we know how to manage our CapEx spend to build out a long-term asset and a lot of the hydration of the kit happens when we have the demand signal. 

There is definitely spend for training. Even there, of course, we will only be scaling training as we see the demand accrue in any given period in time…

…Being able to maybe share a little more about that when we talked about roughly half of FY ’24’s total capital expense as well as half of Q4’s expense, it’s really on land and build and finance leases, and those things really will be monetized over 15 years and beyond. And they’re incredibly flexible because we’ve built a consistent architecture, first with the Commercial Cloud and second with the Azure Stack for AI, regardless of whether the demand is at the platform layer or at the app layer or through third parties and partners or, frankly, our first-party SaaS, it uses the same infrastructure. So it’s a long-lived flexible assets…

…Could we see sort of consistent revenue growth without maybe what you would say is more of this sort of elevated capital expense number or something that continues to accelerate. And the answer to that is yes because there’s 2 different pieces, right? You’re seeing half of this go toward long-term builds that Satya mentioned, the pace at which we fill those builds with CPUs or GPUs will be demand-driven. And so if we see differences in demand signal, we can throttle that investment on the CPU side, which we’ve done for I guess, a long time at this point, as I reflect, and we’ll use all that same learning and demand signal understanding to do the same thing on the GPU side. And so you’re right that you could see relatively consistent revenue patterns and yet see these inconsistencies and capital spend quarter-to-quarter…

…We think about it in terms of what’s the total percentage of cost that goes into each line item, land which obviously has a very different duration and a very different lead time. So those are the other 2 considerations. We think about lead time and duration of the asset. Land, network, construction, the system or the kit and then the ongoing cost. And so if you think about it that way, then you know how to even adjust, if you will, the capital spend based on demand signal.

For Azure’s expected growth of 28%-29% in 2024 Q3 (FY2025 Q1), management expects consumption trends from 2024 Q2 (FY2024 Q4) to continue through FY2025 H1 and the consumption trends include capacity-constrained AI-demand as well as non-AI growth; management expects Azure’s growth to accelerate in FY2025 H2, driven by increase in AI capacity to meet growing demand

 In Azure, we expect Q1 revenue growth to be 28% to 29% in constant currency. Growth will continue to be driven by our consumption business, inclusive of AI, which is growing faster than total Azure. We expect the consumption trends from Q4 to continue through the first half of the year. This includes both AI demand impacted by capacity constraints and non-AI growth trends similar to June. Growth in our per user business will continue to moderate. And in H2, we expect Azure growth to accelerate as our capital investments create an increase in available AI capacity to serve more of the growing demand…

… Capacity constraints, particularly on AI and Azure will remain in Q4 and will remain in H1. 

When Microsoft transitioned to the cloud (in the late 2000s and early 2010s), it was rolled out geography by geography, whereas this current AI platform shift is done globally straight away; Microsoft’s consistent technological infrastructure helps its current AI platform shift achieve faster margin improvement compared to the shift to cloud

You can see what we’re doing and focused on is building out this network in parallel across the globe. Because when we did this last transition, the first transition to the Cloud, which seems a long time ago sometimes. It rolled out quite differently. We rolled out more geo by geo and this one because we have demand on a global basis, we are doing it on a global basis, which is important. We have large customers in every geo… 

…[Question] With Cloud, it took time for margins to improve. It looks like with AI, it’s happening quicker. Can you give us a sense of how you think about the margin impact near term and long term from all the investment on AI?

[Answer] To answer the second half of your question on margin improvement, looking different than it did through the last cloud cycle. That’s primarily for a reason I’ve mentioned a couple of times. We have a consistent platform. So — because we’re building to on Azure AI stack, we don’t have to have multiple infrastructure investments. We’re making one. We’re using that internally first party, and that’s what we’re using with customers to build on as well as ISVs. So it does, in fact, make margins start off better and obviously scale consistently.

Management sees generative AI as fundamentally just being software, and it is translating into growth for Microsoft’s SaaS (software-as-a-service) products; management sees the growth in the usage of Microsoft’s software products as a healthy sign of AI adoption

[Question] How should we think about what it’s going to take for GenAI to become more real across the industry and for it to become more visible within your SaaS offerings?

[Answer] At the end of the day, GenAI is just software. So it is really translating into fundamentally growth on what has been our M365 SaaS offering with a newer offering that is the Copilot SaaS offering, which today is on a growth rate that’s faster than any other previous generation of software we launched as a suite in M365. That’s, I think, the best way to describe it. I mean the numbers I think we shared even this quarter are indicative of this, Mark. So if you look at it, we have both the landing of the seats itself quarter-over-quarter that is growing 60%, right? That’s a pretty good healthy sign. The most healthy sign for me is the fact that customers are coming back there. That is the same customers with whom we landed the seats coming back and buying more seats. And then the number of customers with 10,000-plus seats doubled, right? It’s 2x quarter-over-quarter. That, to me, is a healthy SaaS core business.

Microsoft has dealt with AI capacity constraints by working with third parties who are happy to help Microsoft extend the Azure platform

We’ve talked about now for quite a few quarters, we are constrained on AI capacity. And because of that, actually, we’ve, to your point, have signed up with third parties to help us as we are behind with some leases on AI capacity. We’ve done that with partners who are happy to help us extend the Azure platform, to be able to serve this Azure AI demand. 

Netflix (NASDAQ: NFLX)

Netflix has been using AI (artificial intelligence) and ML (machine learning) for many years to improve the content discovery experience and drive more engagement, and management thinks GenAI (generative AI) has great potential to improve these efforts; but it’s also important ultimately for Netflix to have great content

 We’ve been using similar technologies, AI and ML, for many years to improve the discovery experience and drive more engagement through those improvements. We think that generative AI has tremendous potential to improve our recommendations and discovery systems even further. We want to make it even easier for people to find an amazing story that’s just perfect for them in that moment. But I think it’s also worth noting that the key to our success stacks, right, it’s quality at all levels. So it’s great movies, it’s great TV shows, it’s great games, it’s great live events, and a great and constantly improving recommendation system that helps unlock all of that value for all of those stories.

Management is unsure how AI will specifically impact content creation, but they think AI will result in a great set of creator tools, as there has been a long history of technology improving the content creation process; management thinks that when it comes to content creation, great story-telling is still the most important thing, even as content creators experiment with AI

But I think it’s also worth noting that the key to our success stacks, right, it’s quality at all levels. So it’s great movies, it’s great TV shows, it’s great games, it’s great live events, and a great and constantly improving recommendation system that helps unlock all of that value for all of those stories. nd one thing that’s sure, if you look back over 100 years of entertainment, you can see how great technology and great entertainment work hand in hand to build great, big businesses. You can look no further than animation. Animation didn’t get cheaper, it got better in the move from hand-drawn to CG animation. And more people work in animation today than ever in history. So I’m pretty sure that there’s a better business and a bigger business in making content 10% better than it is making it 50% cheaper…

…I think that shows and movies, they win with the audience when they connect. It’s in the beauty of the writing. It’s in the chemistry of the actors. It’s in the plot, the surprise and the plot twist, all those things…

….So my point is they’re looking to connect. So we have to focus on the quality of the storytelling. There’s a lot of filmmakers and a lot of producers experimenting with AI today. They’re super excited about how useful a tool it can be. And we got to see how that develops before we can make any meaningful predictions on what it means for anybody. But our goal remains unchanged, which is telling great stories.

Nu Holdings (NYSE: NU)

Nu Holdings made a recent acquisition of Hyperlane, a provider of AI solutions in the financial services space; Hyperlane’s AI platform has improved the performance of even Nu Holdings’ most advanced machine learning models when utilising a foundation model focused on financial services that used Nu Holdings’ own unstructured data

I wanted to highlight our recently announced acquisition of Hyperplane. Hyperplane is a Silicon Valley-based leader in AI power solutions for the financial services space. As we tested Hyperplane’s platform on our vast amount of data, we were impressed by the opportunity to meaningfully improve performance of even our most advanced machine learning models by using a financial services focused foundation model that included our own unstructured data. We’re very excited to welcome the Hyperplane team on board and see them as a key part of our AI strategy in the foreseeable future. 

Shopify (NASDAQ: SHOP)

Shopify’s management believes the company can continue to post operating leverage, partly through the internal use of AI to drive productivity

We believe that we can continue to drive operating leverage through 4 key things: disciplined growth in headcount, which we have kept essentially flat for 5 quarters and where we expect we can keep head count growth well below revenue growth; strategic returns-based marketing to support and sustain our long-term revenue growth; internal use of AI and automation to drive productivity; and leveraging and continuing to enhance our internally-built GSD and Shopify OS systems, which allow us to smartly aim the product development work and size the team for maximum impact and efficiency.

Taiwan Semiconductor Manufacturing Company (NYSE: TSM)

TSMC’s capital expenditure is always in anticipation of growth in future years; capex for 2024 is now expected to be US$30 billion to US$32 billion (2023’s capex was US$30.4 billion), up at the low-end from commentary given in the 2024 Q1 earnings call; most of TSMC’s capex are for advanced process technologies; management sees strong structural AI-related demand and is willing to invest to support its customers

Every year, our CapEx is spent in anticipation of the growth that will follow in the future years, and our CapEx and capacity planning is always based on the long-term market demand profile. As the strong structural AI-related demand continues, we continue to invest to support our customers’ growth. We are narrowing the range of our 2024 capital budget to be between USD 30 billion and USD 32 billion as compared to USD 28 million to USD 32 billion previously. Between 70% and 80% of the capital budget will be allocated for advanced process technologies. About 10% to 20% will be spent for specialty technologies, and about 10% will be spent for advanced packaging, testing, mass-making and others. At TSMC, a higher level of capital expenditures is always correlated with the higher growth opportunities in the following years. 

TSMC’s management is seeing a continuation of a strong surge in AI-related demand, which supports structural demand for energy-efficient computing

The continued surge in AI-related demand supports a strong structural demand for energy-efficient computing.

TSMC’s management sees TSMC as a key enabler of AI; management has a disciplined framework, consisting of both a top-down and bottoms-up approach, to plan its capacity buildout; management is not going to make the same kind of mistake it made in 2021 and 2022 when planning its capacity; management has spent a lot of effort studying AI-demand for its capacity-planning and has also asked its customer (likely referring to Nvidia) to be more realistic; management has been testing out AI within TSMC and have found it to be very useful, so management thinks AI demand is real; TSMC has been buying chips from its customer (likely referring to Nvidia)

 As a key enabler of AI applications, the value of our technology position is increasing as customers rely on TSMC to provide the most advanced process and packaging technology at scale in the most efficient and cost-effective manner. As such, TSMC employs a disciplined framework to address the structural increase in the long-term market demand profile underpinned by the industry megatrend of AI, HPC and 5G. We work closely with our customers to plan our capacity. We also have a rigorous and robust system that evaluates and judges market demand from both a top-down and bottom-up approach to determine the appropriate capacity to build…

… [Question] Now looking at GenAI, obviously, the technology has lots of great potential, but a new technology also have lots of volatilities where you start to ramp. And so how are we managing the volatilities of the demand? Why do you think this time around it is different versus COVID period?

[Answer] I thought I explained that our capacity premium process, right, and the investment, we have — I put a wording of discipline. That means we are not going to repeat the same kind of mistake that we have in 2021, 2022. Now this time, again, we look at the overall very big demand forecast for my customer. And so I look at it into actually the whole company with many people now examining and study that really is AI is so used for will be used by a lot of people or not. And we test ourself first inside TSMC, we are using AI, we are using machine learning skill to improve our productivity, and we found out it’s very useful. And so I also in the line to buy my customer’s product, and we have to form in the line, like I cannot privilege here, I’m sorry, but it’s useful.

And so I believe that this time, AI’s demand is more real than 2 or 3 years ago. At that timing it is because people were afraid of a shortage, and so automotive, everything, you name it, they are all in shortage. This time, AI alone only AI alone, it will be a very useful tool for the human being to improve all the productivity in our daily life, be it in medical industry or in any product, manufacturing industry or autonomous driving, everything you need AI. And so I believe it’s more real. But even with that, we also have a top-down bottom-up approach and discuss with our customers and ask them to be more realistic. I don’t want to repeat the same kind of mistake 2 or 3 years ago, and that’s what we are doing right now.

TSMC’s management sees N2, N2P, and A16 as the technologies that will enable TSMC to capture growth opportunities in the years ahead; TSMC’s AI customers are migrating aggressively from N-1 to leading edge nodes, and management is seeing a lot of customers wanting to move into N2, N2P, and A16 quickly, but capacity is very tight and will only loosen in the next year or two years

We believe N2, N2P, N16 and its derivatives will further extend our technology leadership position and enable TSMC to capture the growth opportunities well into the future…

…[Question] We’re hearing that AI chipmakers are looking to migrate more aggressively from N-1 to the leading edge, particularly due to backside power because they’re trying to lower their power budgets going forward. So my question, can you support this move?

[Answer] You are right. All the people want to move into kind of a power-efficient mode. And so they are looking for the more advanced technology so that they can save power consumption. And so a lot of my customers want to move into N2, N2P, A16 quickly. We are working very hard to build the capacity to support them. Today, it’s a little bit tight, not a little bit, actually, today is very tight. I hope in next year or the next 2 years, we can build enough capacity to support this kind of demand. 

TSMC’s management is seeing such high demand for AI-accelerator and CoWoS packaging that supply is so tight; management is hopeful that a balance between demand and supply can be met in 2025 or 2026; it appears that TSMC will be doubling CoWoS capacity again in 2025; CoWoS (or advanced packaging) used to have much lower gross margin than the corporate average, but it is now approaching the corporate average; TSMC is working with its OSAT (outsourced semiconductor assembly and test) partners to expand its CoWoS capacity

[Question] How do you think about supply/demand balance for AI accelerator and CoWoS advanced packaging capacity?

[Answer]  I also tried to reach the supply and demand balance, but I cannot today. The demand is so high. I had to work very hard to meet by customers demand. We continue to increase. I hope sometime in 2025 or 2026, I can reach the balance… The supply continues to be very tight all the way to probably 2025 and hope it can be eased in 2026. That’s today’s situation…

…[Question] Are you going to double your capacity again next year for CoWoS?

[Answer] The last time I say that this year, I doubled it, right, more than double, okay? So next year, if I say double it, probably, I will answer your question again next year, and say more than double, okay? We’re working very hard, as I said, wherever we can, whenever we can…

…For advanced packaging, the gross margin used to be much lower than the corporate average. Now it’s approaching corporate average. We are improving it that’s because of scale of the economics, and we put a lot of effort to reduce our cost. So gross margin is greatly improving in these 2 years…

… I just answered the question whether the CoWoS capacity is enough or not? Is not enough. And in great shortage, and that limited my customers’ growth. So we are working with our OSAT partner and trying to give more capacity to my customer so that they can grow here.

TSMC’s smartphone customers have been using InFO (Integrated Fan-Out) technologies but as they start building edge-AI devices, they are starting to use 3DIC (Three Dimensional Integrated Circuit) and SoIC (System on Integrated Chip) technologies

[Question] In regards to advanced packaging with more and more customers working on edge AI devices without — well, being overly specific, but what does it mean or the implication for advanced packaging solutions that we expect in the next 2 years to see these edge AI customers start to use SoIC or 3DIC particularly smartphone? Will they still be using info? Or will they also consider these solutions as well.

[Answer] As my customer moving into 2-nanometer or A16, they all need to probably take in the approach of chiplets. So once you use your chiplets, you have to use in advanced packaging technologies. On the edge AI, for those kind of smartphone customer, as compared with the HPC customers, HPC is moving faster because of bandwidth concerns, latency of footprint or all those kind of thing. For smartphone customer, they need to pay more attention to the footprint as well as the functionality increase. So you observe my big customers taking the info first and then for a few years, nobody catch it up. They are catching up okay? 

TSMC’s management is seeing a lot of customers wanting to put AI functionality into edge devices; this will increase dye sizes by 5% to 10%, but so far there’s no spike in unit growth of the devices; management thinks the unit growth will happen a few years later as the AI functionalities start to stimulate demand for replacement of older devices

[Question] For silicon content, recall a few years back when 5G just started to ramp you used to provide the silicon content expectations of 5G high-end and mid-end and low-end smartphones, so I wonder at this point of time, if you have any estimates for AI for smartphone going to next 2, 3 years?

[Answer] AI is so hard. So that’s right now everybody — all my customers want to put the AI functionality into the edge devices and so the dye size will be increased, okay? How much? I mean it’s different from my customer-to-customers product. But basically, probably 5% to 10% dye size increase will be a general rule. Unit growth, not yet, okay? Because we did not see kind of unit growth suddenly increased, but we expect this AI functionality was stimulated some of the demand to stimulate the replacement to be shorter. So in terms of unit growth that in a few years later, probably 2 years later, you will start to see a big increase in the edge device that’s a smartphone and the PC.

AI chips have larger die sizes, so TSMC’s management thinks there’s a need to adopt fan-out panel-level packaging eventually, but the technology is currently not mature enough and will need 2-3 years to attain that maturity

[Question] We also see the bigger footprint of the AI chips. So while there are quite some activities about fan-out panel-level packaging. So do you think that, that solution will be mentioned in the mid- to long run? Or does TSMC have any plan to do the related investment?

[Answer] We are looking at this as kind of a panel level fan-out technology. But the maturity today is not yet, so I — personally, I will think it’s about at least 3 years later, okay? In this, within these 3 years, we don’t have any very solid solution for a dye size bigger than 10x of the radical size. Today, we support our customer all the way to 5x, 6x chip size. I’m talking about the [ fuel ] size, the big [indiscernible] size. 2 years later, I believe the panel fan-out will be — start to be introduced and we are working on it.

Tencent (NASDAQ: TCEHY)

Tencent’s advertising business is benefitting from better click through rates driven by AI; management sees AI technology increasing advertising conversion rates by 10%

We are benefiting from deployment of neural network artificial intelligence on a GPU infrastructure to boost the click-through rate on our advertising inventory…

…And at the same time, on the ad recommendation end, if we can actually increase conversion by 10%, right, that’s sort of pretty modest improvement. The revenue actually grows quite a bit, right? So I think that’s areas in which we are leveraging AI to deliver material and tangible commercial results.

Tencent’s AI-related external revenue is growing, and the company recently launched 3 AI-powered solutions for enterprises, namely image generation engine, video generation engine, and knowledge engine

Tencent Meeting deepened its adoption and monetization, especially in the pharmaceutical manufacturing and retail sectors. We’re generating increasing AI-related external revenue from customers utilizing our high-performance computing infrastructure, such as GPUs and our model library services. We’re generating increasing AI-related external revenue from customers utilizing our high-performance computing infrastructure, such as GPUs and our model library services. We recently launched 3 AI-powered platform solutions for enterprises, image generation engine and video generation engine, which are pretty useful for advertisers creating ad content; as well as knowledge engine, which is particularly useful for finance, education and retail-related services, deploying customer service chat bots.

Tencent’s operating capex in 2024 Q2 was up 144% year-on-year because of investments in GPUs and CPUs; non-operating capex was up 53% year-on-year, driven by construction, but down 80% sequentially

Operating CapEx was RMB 7.2 billion, up 144% year-on-year driven by investment in GPU and CPU servers.  Non-operating CapEx was RMB 1.5 billion, up 53% year-on-year, driven by construction and progress. On a quarter-on-quarter basis, non-operating CapEx was down 80% from the high base in the prior quarter. As a result, total CapEx was RMB 8.7 billion, up 121% year-on-year.

Tencent’s management thinks of AI as more than just large language models

We look at AI as a more complete suite than just large language model. There are the neural networks, machine learning-based recommendation engines, which we use for content recommendation, video recommendation as well as the talking in the ads and content use case, which is already delivering very good result.

Tencent has delivered better content to users through the use of AI

If you take Video Accounts as an example, by using AI, we actually are able to deliver better content and that generates more use time — a pretty big part of the growth in terms of the Video Accounts user time. It’s actually driven by better targeting, better recommendation and that’s in turn driven by AI.

Tencent’s management thinks AI can improve PVE (player vs environment) games by making the computer smarter

In the area of games, we’re actually using AI to bridge the gap between PVE and PVP, right? So when you have games, which allow people to play against other players, but at the same time, sometimes you actually want to create a game mode in which a player actually play against the machine, right? Then — in the past, the machine is actually quite dumb, right? And with AI, we can actually make the machine play like a real player. And we can actually sort of have it to play a varying levels of skills and make the user experience and the gameplay very fun.

Tencent’s management’s focus with LLMs is to improve the technology; Tencent has already built a MOE (mixture of experts) architecture model, which is one of the top AI models in the Chinese language; Tencent is deploying its LLM in Yuanbao, an app launched to allow users to interact with its LLM; Tencent’s LLM is improving search results and Yuanbao is getting positive feedback; when Yuanbao improves, management will increase promotional resources to increase the user base; management also wants to incorporate Yuanbao into different parts of its ecosystem

Now in terms of LLM, the key thing for us is actually improving the technology. And as we shared before, we have already built an MOE architecture model, which is performing as one of the top models in China. And when compared with international models on Chinese language, I think we are at that top of the pack. And we are deploying our LLM in Yuanbao, which is an app that we have launched which allowed users to interact with our large language model in multiple ways. And one way is enhanced search functionality so that users can actually ask a question. And based on search results, we can actually provide a very direct answer to the questions that our users pose and we have rolled it out to a large enough sample size to get user feedback and the feedback so far has been quite positive…

…Over time, Yuanbao, when it gets to a certain level of quality, then we’re going to increase our promotional resources and try to get more users into the app. And at the same time, when it gets to an even better level of expertise, then we can actually start incorporating it into different parts of our ecosystem. We have a lot of apps which actually has got interaction use cases, which we can leverage our generative AI technology.

Renting out GPUs for AI workloads is a big business in China too, but it’s to a smaller extent when compared to what’s happening in the USA; Tencent’s management is seeing very fast growth in demand for GPU-rentals for AI needs partly because the growth is happening off a low base; the demand for GPU-rentals is partially cannibalising the demand for CPUs

Clearly, for the U.S. hyperscale Cloud providers, renting out GPUs to other companies with AI requirements has become a very big business. In China, the same trend is evident, but to a lesser extent because you don’t have the same multitude of extremely well-funded start-ups trying to build large language models on their own in China. There are many small companies, but they’re capitalized for $1 billion, $2 billion. They’re not capitalized at $10 billion or $90 billion, other way that some of the giant U.S. VC-funded start-ups are now capitalized in the space. And it’s also a somewhat challenging economic environment. Now that said, we have seen that within our Cloud, the demand from customers for renting GPUs for their own AI needs has been growing very swiftly. The percentage growth rates are very fast, but they’re very fast partly because it’s a low base. And also partly because, while some of that demand for renting GPUs in the Cloud is incremental, some of it is replacing demands that would otherwise have existed anyway for renting CPUs in the Cloud. And so while the business of GPU provision is doing very well, the business of CPU processing is more flat because the incremental demand is for GPU, not CPU.

Tesla (NASDAQ: TSLA)

Tesla has made a lot of progress with full self-driving in Q2; a new version, version 12.5, of the autonomous software has just started to be rolled out; version 12.5 of the FSD (full self-driving) software is a step-change improvement in supervised full self-driving; management thinks that most people still do not know how good version 12.5 is; as Tesla increases the miles between intervention, the system can transition from supervised full self-driving to unsupervised full self-driving; management would be shocked if Tesla cannot achieve unsupervised full self-driving next year, but they also note that they have been overly optimistic on the timeline for self-driving; management believes that Tesla will be able to get regulatory approval for unsupervised full self-driving once it shows the rate of accidents is less than human driving; self-driving capabilities of Tesla vehicles outside of North America are far behind those of Tesla vehicles in North America; management is asking for regulatory approval of Tesla supervised full self-driving in Europe, China, and other countries, and the approvals, which are expected before end-2024, will be a driver of demand for Tesla vehicles; FSD uptake is still low despite some increase after a recent price reduction

Regarding full self-driving and Robotaxi, we’ve made a lot of progress with full self-driving in Q2. And with version 12.5 beginning rollout, we think customers will experience a step change improvement in how well supervised full self-driving works. Version 12.5 has 5x the parameters of 12.4 and finally merged the highway and city stacks. So the highway stack at this point is pretty old. So often the issues people encounter are on the highway. But with 12.5, we finally merged the 2 stacks. I still find that most people actually don’t know how good the system is. And I would encourage anyone to understand the system better to simply try it out and let the car drive you around…

…And as we increase the miles between intervention, it will transition from supervised full self-driving to unsupervised full self-driving, and we can unlock massive potential [ in the fleet ]…

…I guess that, that’s really just a question of when can we expect the first — or when can we do unsupervised full self-driving. It’s difficult, obviously, my predictions on this have been overly optimistic in the past. So I mean, based on the current trend, it seems as though we should get miles between interventions to be high enough that — to be far enough in excess of humans that you could do unsupervised possibly by the end of this year. I would be shocked if we cannot do it next year. So next year seems highly probable to me based on quite simply plus the points of the curve of miles between intervention. That trend exceeds the humans for sure next year, so yes…

…So it’s this capability. I think in our experience, once we demonstrate that something is safe enough or significantly safer than human, we find that regulators are supportive of deployment of that capability. It’s difficult to argue with — if you have got a large number of — if you’ve got billions of miles that show that in the future unsupervised FSD is safer than human, what regulator could really stand in the way of that. They’re morally obligated to approve. So I don’t think regulatory approval will be a limiting factor. I should also say that the self-driving capabilities that are deployed outside of North America are far behind that in North America. So with Version 12.5, and maybe 12.6, but pretty soon, we will ask for regulatory approval of the Tesla supervised FSD in Europe, China and other countries. And I think we’re likely to receive that before the end of the year. There will be a helpful demand driver in those regions…

…[Question] You mentioned that FSD take rates were up materially after you reduced the price. Is there any way you can help us quantify what that means exactly?

[Answer] We’ve shared that how — that we’ve seen a meaningful increase. I don’t want to get into specifics because we started from a low base, but we are seeing encouraging results. 

Tesla will unveil its robotaxi product on 10th of October, after postponing it for a few months; the current plan is for robotaxis to be produced in Tesla’s headquarters at Giga Texas; management’s aim is to have a robotaxi fleet that’s made up of both Tesla-owned vehicles and consumer-owned vehicles, and consumers can rent out their cars, just like renting out their apartments for Airbnb; Tesla has a clause with every vehicle purchase that Tesla vehicles can only be used in the Tesla fleet and not in any 3rd-party autonomy fleet; management believes that once unsupervised full self-driving is available, most people will rent out their Tesla vehicles, so the Tesla robotaxi service will achieve instant scale given the existing number of Teslas on the road

We postponed the sort of robotaxi product unveil by a couple of months where it’s shifted to 10/10, to the 10th of October. And this is because I wanted to make some important changes that I think would improve the vehicle — the sort of — the Robotaxi — the thing — the main thing that we’re going to show…

…And I should say that the Cybertaxi or Robotaxi will be locally produced here at our headquarters at Giga Texas… 

…This would just be the Tesla network. You just literally open the Tesla app and summon a car and we send a car to pick you up and take you somewhere. And our — we will have a fleet that’s on the order of 7 million [ vehicle autonomy ] soon. In the U.S. it will be over 10 million and over 20 million. This is in that scale. And the car is able to operate 24/7 unlike the human drivers. So the capability to — like this basically instant scale with a software update. And now this is for a customer-owned fleet. So you can think of that as being a bit like Airbnb, like you can choose to allow your car to be used by the fleet or cancel that and bring it back. It will be used by the fleet all the time, can be used by the fleet some of the time and then Tesla will take a share in the revenue with the customer…

…And there’s an important clause we’ve put in every Tesla purchase, which is that the Tesla vehicles can only be used in the Tesla fleet. They cannot be used by a third party for autonomy…

…[Question] Do you think that scales like progressively, so you can start in a city with just a handful of cars. Then you grow the number of cars over time? Or do you think there is like a critical mass you need to get to, to be able to offer like a service that is of competitive quality compared to what like Uber would be typically delivering already?

[Answer] I guess I’m not — I’m not conveying this correctly. The entire Tesla fleet basically becomes active. This is obviously — maybe there’s some number of people who don’t want their car to earn money. But I think most people will. It’s instant scale.

Tesla is nearing completion of the South expansion of Giga Texas, which is Tesla’s largest training cluster of GPUs to-date; there was a story earlier this year that Tesla sent its new H100 AI chip deliveries to Elon Musk’s other entities but this happened only because Tesla had no place to house the chips at that point in time; Tesla now has a place for the chips because of the South expansion of Giga Texas

We’re also nearing completion of the South expansion of Giga Texas, which will house our largest training cluster to date. So it will be an incremental 50,000 H100s, plus 20,000 of our hardware for AI5, Tesla AI computer…

…I mean I think you’re referring to a very — like an old article regarding GPUs. I think that’s like 6 or 7 months old. Tesla simply had no place to turn them on. So it would have been a waste of Tesla Capital because we would just have to order H100s and have no place to turn them on. So I was just – this wasn’t a let’s pick xAI over Tesla. There was no — the Tesla test centers were full. There was no place to actually put them. The — we’ve been working 24/7 to complete the South extension on the Tesla [indiscernible] Texas. That self extension is what will house the 50,000 H100s, and we’re beginning to move the certain H100 server racks in place there. But we really needed — we needed that to complete basically. You can’t just order compute — order GPUs and turn them on, you need a data center. So I want to be clear, that was in Tesla’s interest, not contrary to Tesla’s interest. Does Tesla no good to have GPUs that it can’t turn on. That South extension is able to take GPUs, which is really just this week. We are moving the GPUs in there and we’ll bring them online.

The Optimus robot is already performing tasks in Tesla’s factory; management expects to start limited production of Optimus in early 2025; early production is for Tesla’s consumption, and management expects a few thousand robots in Tesla’s factories by end-2025; management expects Optimus to enter high-volume production in 2026 and to release Optimus to external customers by then; management believes that Optimus will be the biggest revenue contributor to Tesla in the future, with an estimated total addressable market of 20 billion units of Optimus robots; management thinks Tesla has all the ingredients to build large scale, generalised humanoid robots 

With Optimus, Optimus is already performing tasks in our factory. And we expect to have Optimist production Version 1 and limited production starting early next year. This will be for Tesla consumption. It’s just better for us to iron out the issues ourselves. But we expect to have several thousand Optimus robots produced and doing useful things by the end of next year in the Tesla factories. And then in 2026, ramping up production quite a bit. And at that point, we’ll be providing Optimus robots to outside customers. There will be a production Version 2 of Optimus…

…I mean, as I said a few times, I think the long-term value of Optimus will exceed that of everything else that Tesla combined. So it’s simply just never considered the usefulness, utility of a humanoid robot that can do pretty much anything you asked of it. II think everyone on earth is going to want one. There are 8 billion people on earth. So it’s 8 billion right there. Then you’ve got all of the industrial uses, which is probably at least as much, if not, way more. So I suspect that the long term demand for general purpose humanoid robots is in excess of 20 billion units. And Tesla has the most advanced humanoid robot in the world and is also very good at manufacturing, which these other companies are not. And we’ve got a lot of experience with — the most experienced — we’re the word leaders in [ Real World AI ]. So we have all of the ingredients. I think we’re unique in having all of the ingredients necessary for large scale, high utility, generalized humanoid robots.

Management expects capex to be over US$10 billion in 2024 (was US$8.9 billion in 2023) because of spending on the AI GPU cluster

On the CapEx front, while we saw a sequential decline in Q2, we still expect the year to be over $10 billion in CapEx as we increase our spend to bring a 50 GPU cluster on luck. This new center will immensely increase our capabilities to scale FSD and other AI initiatives. 

Tesla will continue working on its own AI GPU called Dojo to reduce reliance on NVIDIA, and also because NVIDIA’s supply for GPUs is so tight; management sees a path where Dojo’s chips can be competitive with NVIDIA’s

So Dojo, I should preface this by saying I’m incredibly impressed by NVIDIA’s execution and the capability of their hardware. And what we are seeing is that the demand for NVIDIA hardware is so high that it’s often difficult to get the GPUs. And there just seems this — I guess I’m quite concerned about actually being able to get steady out NVIDIA GPUs and when we want them. And I think this therefore requires that we put a lot more effort on Dojo in order to have — in order to ensure that we’ve got the training capability that we need. So we are going to double down on Dojo and we do see a path to being competitive with NVIDIA with Dojo. And I think we kind of have no choice because the demand for NVIDIA is so high and it’s obviously their obligation essentially to raise the price of GPUs to whatever the market will bear, which is very high. So I think we’ve really got to make Dojo work and we will.

Tesla is learning from Elon Musk’s AI startup, xAI; Musk is aware that Tesla needs shareholder approval before the company can invest in xAI, but he thinks it’s a good idea; Musk sees opportunities to integrate xAI’s foundation model, Grok, into Tesla’s software; Musk found that some engineers are only interested in working on AGI (artificial general intelligence) and they would have gone to other AI startups if Musk was not working on xAI since they would not have chosen Tesla anyway

Tesla is learning quite a bit from xAI. It’s been actually helpful in advancing full self-driving and in building up the new Tesla data center. With — regarding investing in xAI, I think, we need to have a shareholder approval of any such investment. But I’m certainly supportive of that if shareholders are, the group — probably, I think we need a vote on that. And I think there are opportunities to integrate Grok into Tesla’s software, yes…

…With regard to xAI, there are a few that only want to work on AGI. So what I was finding was that when trying to recruit people to Tesla, they were only interested in working on AGI and not on Tesla’s specific problems and they want to start — do a start-up. So it was a case of either they go to a startup or — and I am involved or they do a start-up and I am not involved. Those are the 2 choices. This wasn’t they would come to Tesla. They were not going to come to Tesla under any circumstances…

…I tried to recruit them to Tesla, including to say, like, you can work on AGI if you want and they refused. Only then was xAI created.

Management still thinks Tesla can rent out latent AI inferencing compute for general computing purposes from its fleet of vehicles (and perhaps humanoid robots) in the future

Just distributed compute. It seems like a pretty obvious thing to do. I think where the distributed compute becomes interesting is with next-generation Tesla AI truck, which is hardware viable, what we’re calling AI5, which is from the standpoint of inference capability comparable to B200 and [ a bit of ] B200. And we’re aiming to have that in production at the end of next year and scale production in ’26. So it just seems like if you’ve got autonomous vehicles that are operating for 50 or 60 hours a week, there’s 168 hours in a week. So we have somewhere above, I think, 100 neural net computing. I think we need a better word than GPU because GPU means graphics processing unit. So there’s a 100 hours plus per week of AI compute, AI [ first ] compute from the fleet in the vehicles and probably some percentage from humanoid robots. That it would make sense to do distributed inference. And if there’s a fleet of at some point, 100 million vehicles with AI5 and beyond, AI6 and 7 and what not and there are maybe billions of humanoid robots. That is just a staggering amount of inference compute that could be used for general purpose computing. Doesn’t have to use it for the humanoid robot or for the car.

Management believes that Waymo’s approach to autonomous vehicles is a localised solution that requires high-density mapping and is thus quite fragile compared to Tesla’s approach

I mean our solution is a generalized solution like what everybody else has. You could see if Waymo has [ one of it ], they have very localized solution that requires high-density mapping. It’s not — it’s quite fragile. So their ability to expand, I believe, is limited. Our solution is a general solution that works anywhere. It would even work on a different earth. So if you [ branded ] a new earth, it would work on new earth…

…in terms of regulatory approval, the vehicles are governed by FMVSS in U.S., which is the same across all 50 states. The road rules are the same across all 50 states. So creating a generalized solution gives us the best opportunity to deploy in all 50 states reasonably. Of course, there are state and even local municipal level regulations that may apply to being a transportation company or deploying taxis. But as far as getting the vehicle on the road, that’s all federal and that’s very much in line with what Elon was suggesting of the data and the vehicle itself…

…To add to the technology point, the end-to-end network basically makes no assumption about the location. Like you could add data from different countries and it just like performs equally well there. That’s like almost close to 0, U.S. specific code in there. It’s all just the data that comes from the U.S.

Visa (NYSE: V)

Visa’s management is investing in AI, particularly generative AI (genAI), because the company has use-cases for the technology in areas such as fraud reduction and productivity improvement; management is very optimistic about the positive impact that generative AI can have 

First of all, to frame it is we are all in on GenAI at Visa as we’ve been all in on predictive AI for more than a decade. We’re applying it in 2 broad-based different ways. One is sort of adopting across the company to drive productivity and we’re seeing real results there. We’re seeing great results, great adoption, great productivity increases from technology to accounting to sales all across the company. The second is applying generative AI to enhance the entire payment ecosystem. And to the latter part of your question, absolutely. I guess I’d give you one set of examples or some of the risk tools and capabilities that we’ve been deploying in the market. I mentioned the risk products that we’re using on RTP and account-to-account payments. That is an opportunity to reduce fraud, both for merchants and for issuers. I think I mentioned on a previous call, we have our Visa Provisioning Intelligence Service, which is using artificial intelligence to help predict token provisioning fraud before it happens. That also is a benefit to both issuers and merchants. And the list goes on. So we are very optimistic about the positive impact that generative AI can have, not just on our own productivity but on our ability to help drive increased sales and lower fraud across the ecosystem.

Wix (NASDAQ: WIX)

Wix’s management continues to improve the company’s AI capabilities; Wix has released 17 AI business assistants to-date; the AI business assistants support a wide range of use cases and Wix has already received positive feedback on them; Wix will be releasing dozens more AI assistant later in 2024; the 17 business assistants are all customer-facing but the assistants can play one of two roles, (1) be a question-and-answer AI assistant, and (2) be an assistant that executes actions; the AI business assistants rarely hallucinate; management wants to add these AI assistants everywhere in the Wix product suite

We continue to build up our suite of AI capabilities as a result of the numerous AI initiatives and work streams across Wix. Last quarter, we introduced our plan to embed AI assistance across our platform and products. I’m excited to share that we have released 17 AI business assistants so far to date. These assistants span a wide range of use cases to support users with minimal hands-on support, thus streamlining their experience. These conversational AI assistants act as a right-hand aid for users to guide them through the entire life cycle of ideating, creating and managing their online presence. Our offering includes an analytics assistant that can help Wix users find the data they need without having to search through dozens of reports, and an assistant that helps users create events through a conversational chat. We have already received positive feedback on this first set of AI assistants with dozens more set to launch later this year…

…how many of the 17 are customer-facing? And the answer is all of them. The concept is that we are currently — we build a platform in which it is easier for us to build an AI assistant. And then that enable us to develop 2 kinds of different assistants. The first one would be a question-and-answer AI assistant, so if you have a product like booking, how do I add a staff member to my yoga studio, right? And so you can actually talk to the AI and ask questions, get answered, and ask question, get answer, as you would do with the normal human being. And then we see a great result in that in terms of how customers quickly find the answers. Hallucinations are very small, the percentage, probably similar to what a human would do or not even better…

…The other thing that we are doing is that you can ask questions and you can have the AI do things for you. So this is the second kind. And for example, if you go to our analytics, you see that you can actually start asking questions and get the reports done for you automatically by the AI. So this is an AI that activates other agents in order to give you answers or do actions for you. How do I make an event that is a wedding event? What not? And then it will do — analyze [ VP ]. But if you want to create an event which is selling tickets for a concert, it will define that, willing to work with you on that. So those kind of things streamline and reduce a lot of friction from the customer…

…We’re going to add those kind of assistants in pretty much everywhere that we can on Wix. 

Wix’s management launched AI creation capabilities for its mobile app builder in June 2024, which enables users to create and edit iOS and Android apps through a cha 

We launched AI creation capabilities for our mobile app builder in June. This new solution enables users to create and edit iOS or Android apps through an AI chat experience. Once AI understands the user’s goals, intent and desired aesthetic, our technology generates a branded app that can be customized and managed from the App Editor.

Wix’s management recently released new AI features to help users with content-generation

We also recently released a suite of new AI features designed to help users identify relevant topics for blogs as well as generate outlined content and images for their target audience. With this new experience, users can swiftly turn ideas into new ready articles, significantly reducing the time and effort required to create engaging content, and ultimately, changing the blog creation experience.

Wix’s management sees both Self Creators and Partners having excellent engagement with Wix’s AI tools; management expects Wix’s AI tools to be a competitive advantage and significant driver of future growth; Wix’s AI tools continue to drive user conversion; Wix released its first AI product all the way back in 2016 and management saw that the AI functionality had very high adoption and drove dramatic improvement in user conversion; the latest version of the AI product, released earlier this year, had the same effect; Wix’s AI agents are having measurable positive impact on engagement; management thinks that their 7-8 years of experience with releasing AI technology is helping them integrate AI into Wix’s product suite in a highly intuitive way

Both Self Creators and Partners continue to show excellent engagement with our AI tools. As we expand the breadth of our AI technology, we expect it to continue to be a competitive advantage for us as well as a significant driver of growth going forward…

… Our AI tools continue to drive user conversion…

…Released ADI, the first AI product — GenAI product, actually created website right in the end of 2016. And since then, we’ve seen that by exposing users to AI functionality as part of the natural progression in the product life cycle, we get very high adoption, obviously using those kind of tools and results that can improve. And for ADI, we show that we improved the conversion dramatically. The new version that came earlier this year did it again. And we are seeing that a lot of the agents that we have now, AI agents, when they start to pick up more user interactions and more user conversations, again, create measurable effect. So I’m very optimistic. I think that our experience in releasing AI technology, right, which is almost, what, 8 years now — 7 years now, is helping us understand how to integrate them into the product in a way that actually mixed user interact with them and that they feel natural and don’t feel like you’re stepping out of what you’re doing to do something else and then coming back. And I think that creates a big difference. So yes, I’m very optimistic on the potential that we’re going to see a continuation of the improvement.

There is a big difference between what an agency and a Self Creators need from AI. So for me, if I want to design a website, and I’m not a designer, I want AI to help me design it because English is not my first language and I’m not writing so well in Hebrew as well, right? So I would love AI to also help me write great text and generate images.

When you’re an agency, you probably know how to design and you have your system of design and how things should look like. So you don’t need that. You probably need a little bit to help with the text, but other things, like the image editing, right, and the content recomposition create tremendous value. And then the other things that — in addition to that, for example, a great designer not necessarily know how to configure things to work in a responsive way on different screen resolutions, and we have an AI to do that. So we are utilizing those kind of technologies to streamline the agency’s experience and work and efficiency in a way that is significant to them. I think we have some ideas on how to make it even more significant going forward.

Wix’s management thinks there’s a long way to go before AI technology will make agencies become obsolete by having the computer know automatically what website you want to build and get it fully functioning, so agencies will still be an important business for Wix for many years

In theory, if you can just one day talk to a computer and get the full website functioning that knows exactly what should be there and that it’s easy to update then maybe some of the agency’s business will disappear. But there is a long way until we get to something similar to that. And I think the majority of businesses in the case that they need a website, they want somebody to be responsible for it, somebody that know how to activate the tools and use them and utilize them, and that’s why they go to agencies because they have a professional that understand how to take care of all of their business needs. And there’s a lot of those, right from SEO to how do you write things correctly in order to get the right shipping rules, and there’s a ton of things. So I think that where there’s a long way for AI to go before it can successfully replace good agencies. 

Unless, of course, you are a self-creator by nature, which is a lot of most of our customers, and you want to create your website, you can control it and you can do those things and you can change it. So I think the difference is in the user type and user intent and not necessarily in technology, which I believe means that both will continue to grow, agencies and Self Creators.

Wix’s management is seeing that the newer users who join Wix are those who use more AI tools to automate website creation as compared to earlier users; the presence of Wix’s AI tools opens up new types of customers for Wix

One of the qualification that you needed to have in order to be able to use Wix in the past was to know how to design to some level, to know how to write text to some level and to trust yourself that you’re good enough to do it, right? And then — so most of our users feel that they know how to do those things. And naturally, they will use less AI because they think they can just do it. And I think we are now opening to users that don’t feel that, right? They don’t expect themselves to know how to do those things and expect us to have the tools to — AI tools to automate it for them. So we are already seeing some of this gap, and I believe that this will continue to grow. And essentially, we are opening Wix to be more useful to more new types of customers.

---

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Alphabet, Amazon, Apple, ASML, Coupang, Datadog, Fiverr, Mastercard, Meta Platforms, Microsoft, Netflix, Nu Holdings, Shopify, TSMC, Tesla, Visa, and Wix. Holdings are subject to change at any time.

JPMorgan Chase (NYSE: JPM) is currently the largest bank in the USA by total assets. Because of this status, JPMorgan is naturally able to feel the pulse of the country’s economy. The bank’s latest earnings conference call – for the second quarter of 2024 – was held three weeks ago and contained useful insights on the state of American consumers and businesses. The bottom-line is this: The US economy is stronger than what many would have thought a few years ago given the current monetary conditions, but there are signs of weakness such as slightly higher unemployment and slower GDP growth; at the same time,  inflation and interest rates may stay higher than the market expects, and the Fed’s quantitative tightening may have unpredictable consequences.

What’s shown between the two horizontal lines below are quotes from JPMorgan’s management team that I picked up from the call.

---

1. Broader financial market conditions suggest a benign economic outlook, but JPMorgan’s management continue to be vigilant about potential tail risks; management is concerned about inflation and interest rates staying higher than the market expects, and the effects of the Federal Reserve’s quantitative tightening

While market valuations and credit spreads seem to reflect a rather benign economic outlook, we continue to be vigilant about potential tail risks. These tail risks are the same ones that we have mentioned before. The geopolitical situation remains complex and potentially the most dangerous since World War II — though its outcome and effect on the global economy remain unknown. Next, there has been some progress bringing inflation down, but there are still multiple inflationary forces in front of us: large fiscal deficits, infrastructure needs, restructuring of trade and remilitarization of the world. Therefore, inflation and interest rates may stay higher than the market expects. And finally, we still do not know the full effects of quantitative tightening on this scale.

2. Net charge-offs (effectively bad loans that JPMorgan can’t recover) rose from US$1.4 billion a year ago, mostly because of card-related credit losses that are normalising to historical norms

Credit costs were $3.1 billion, reflecting net charge-offs of $2.2 billion and a net reserve build of $821 million. Net charge-offs were up $820 million year-on-year, predominantly driven by Card…

…I still feel like when it comes to Card charge-offs and delinquencies, there’s just not much to see there. It’s still — it’s normalization, not deterioration. It’s in line with expectations. 

3. JPMorgan’s credit card outstanding loans was up double-digits

Card outstandings were up 12% due to strong account acquisition and the continued normalization of revolve.

4. Auto originations are down

In auto, originations were $10.8 billion, down 10%, coming off strong originations from a year ago while continuing to maintain healthy margins. 

5. JPMorgan’s investment banking fees had strong growth in 2024 Q2, partly because of favourable market conditions; management is cautiously optimistic about the level of appetite that companies have for capital markets activity, but headwinds persist 

This quarter, IB fees were up 50% year-on-year, and we ranked #1, with year-to-date wallet share of 9.5%. In advisory, fees were up 45% primarily driven by the closing of a few large deals and a weak prior year quarter. Underwriting fees were up meaningfully, with equity up 56% and debt up 51%, benefiting from favorable market conditions. In terms of the outlook, we’re pleased with both the year-on-year and sequential improvement in the quarter. We remain cautiously optimistic about the pipeline, although many of the same headwinds are still in effect. It’s also worth noting that pull-forward refinancing activity was a meaningful contributor to the strong performance in the first half of the year…

…In terms of dialogue and engagement, it’s definitely elevated. So I would say the dialogue on ECM [Equity Capital Markets] s elevated and the dialogue on M&A is quite robust as well. So all of those are good things that encourage us and make us hopeful that we could be seeing sort of a better trend in this space. But there are some important caveats.

So on the DCM [Debt Capital Markets] side, yes, we made pull-forward comments in the first quarter, but we still feel that this second quarter still reflects a bunch of pull-forward, and therefore, we’re reasonably cautious about the second half of the year. Importantly, a lot of the activity is refinancing activity as opposed to, for example, acquisition finance. So the fact that M&A remains still relatively muted in terms of actual deals has knock-on effects on DCM as well. And when a higher percentage of the wallet is refi-ed, then the pull-forward risk becomes a little bit higher.

On ECM, if you look at it kind of [ at a removed ], you might ask the question, given the performance of the overall indices, you would think it would be a really booming environment for IPOs, for example. And while it’s improving, it’s not quite as good as you would otherwise expect. And that’s driven by a variety of factors, including the fact that, as has been widely discussed, that extent to which the performance of the large industries is driven by like a few stocks, the sort of mid-cap tech growth space and other spaces that would typically be driving IPOs have had much more muted performance. Also, a lot of the private capital that was raised a couple of years ago was raised at pretty high valuations. And so in some cases, people looking at IPOs could be looking at down rounds, that’s an issue. And while secondary market performance of IPOs has improved meaningfully, in some cases, people still have concerns about that. So those are a little bit of overhang on that space. I think we can hope that over time that fades away and the trend gets a bit more robust.

And yes, on the advisory side, the regulatory overhang is there, remains there. And so we’ll just have to see how that plays out.

6. Management is seeing muted demand for new loans from companies as current economic conditions make them cautious

Demand for new loans remains muted as middle market and large corporate clients remain somewhat cautious due to the economic environment and revolver utilization continues to be below pre-pandemic levels. 

7. Demand for loans in the commercial real estate (CRE) market is muted

In CRE, higher rates continue to suppress both loan origination and payoff activity.

8. Lower income cohorts are facing a little more pressure than higher income cohorts because even though the US economy is stronger than what many would have thought a few years ago given the current monetary conditions, there is currently slightly higher unemployment and slower GDP growth

As I say, we always look quite closely inside the cohort, inside the income cohorts. And when you look in there, specifically, for example, on spend patterns, you can see a little bit of evidence of behavior that’s consistent with a little bit of weakness in the lower-income segments, where you see a little bit of rotation of the spend out of discretionary into nondiscretionary. But the effects are really quite subtle, and in my mind, definitely entirely consistent with the type of economic environment that we’re seeing, which, while very strong and certainly a lot stronger than anyone would have thought given the tightness of monetary conditions, say, like they’ve been predicting it a couple of years ago or whatever, you are seeing slightly higher unemployment, you are seeing moderating GDP growth. And so it’s not entirely surprising that you’re seeing a tiny bit of weakness in some pockets of spend. 

9. The increase in nonaccrual loans in the Corporate & Investment Bank business is not a broader sign of cracks happening in the business

[Question] I know your numbers are still quite low, but in the Corporate & Investment Bank, you had about a $500 million pickup in nonaccrual loans. Can you share with us what are you seeing in C&I? Are there any early signs of cracks or anything?

[Answer] I think the short answer is no, we’re not really seeing early signs of cracks in C&I. I mean, yes, I agree with you like the C&I charge-off rate has been very, very low for a long time. I think we emphasized that at last year’s Investor Day. If I remember correctly, I think the C&I charge-off rate [ over the preceding ] 10 years was something like literally 0. So that is clearly very low by historical standards. And while we take a lot of pride in that number and I think it reflects the discipline in our underwriting process and the strength of our credit culture across bankers and the risk team, that’s not — we don’t actually run that franchise to like a 0 loss expectation. So you have to assume there will be some upward pressure on that. But in any given quarter, the C&I numbers tend to be quite lumpy and quite idiosyncratic. So I don’t think that anything in the current quarter’s results is indicative of anything broader and I haven’t heard anyone internally talk that way, I would say.

10. Management is unwilling to lower their standards for risk-taking just because it has excess capital because they think it makes sense to be patient now given their current assessment of economic risk

And of course, for the rest of the loan space, the last thing that we’re going to do is have the excess capital mean that we lean in to lending that is not inside our risk appetite or inside our credit box, especially in a world where spreads are quite compressed and terms are under pressure. So there’s always a balance between capital deployment and assessing economic risk rationally. And frankly, that is, in some sense, a microcosm of the larger challenge that we have right now. When I talked about if there was ever a moment where the opportunity cost of not deploying the capital relative to how attractive the opportunities outside the walls of the company are, now would be it in terms of being patient. That’s a little bit one example of what I was referring to.

---

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I don’t have a vested interest in any company mentioned. Holdings are subject to change at any time.

Note: This article is a copy of Why It’s So Difficult To Tell When The Stock Market Will Peak that I published more than four years ago on 21 February 2020. With the US stock market at new all-time highs, I thought it would be great to revisit this piece. The content in the paragraphs and table near the end of the article have been revised to include the latest valuation and returns data. 

Here’s a common misconception I’ve noticed that investors have about the stock market: They think that it’s easy to figure out when stocks will hit a peak. Unfortunately, that’s not an easy task at all.

In a 2017 Bloomberg article, investor Ben Carlson showed the level of various financial data that were found at the start of each of the 15 bear markets that US stocks have experienced since World War II:

The financial data that Carlson presented include valuations for US stocks (the trailing P/E ratio,  the cyclically adjusted P/E ratio, and the dividend yield), interest rates (the 10 year treasury yield), and the inflation rate. These are major things that the financial media and many investors pay attention to. (The cyclically-adjusted P/E ratio is calculated by dividing a stock’s price with the 10-year average of its inflation-adjusted earnings.)

But these numbers are not useful in helping us determine when stocks will peak. Bear markets have started when valuations, interest rates, and inflation were high as well as low. This is why it’s so tough to tell when stocks will fall. 

None of the above is meant to say that we should ignore valuations or other important financial data. For instance, the starting valuation for stocks does have a heavy say on their eventual long-term return. This is shown in the chart below. It uses data from economist Robert Shiller on the S&P 500 from 1871 to June 2024 and shows the returns of the index against its starting valuation for 10-year holding periods. It’s clear that the S&P 500 has historically produced higher returns when it was cheap compared to when it was expensive.

But even then, the dispersion in 10-year returns for the S&P 500 can be huge for a given valuation level. Right now, the S&P 500 has a cyclically-adjusted P/E ratio of around 35. The table below shows the 10-year annual returns that the index has historically produced whenever it had a CAPE ratio of more than 30.

If it’s so hard for us to tell when bear markets will occur, what can we do as investors? It’s simple: We can stay invested. Despite the occurrence of numerous bear markets since World War II, the US stock market has still increased by 532,413% (after dividends) from 1945 to June 2024. That’s a solid return of 11.4% per year. Yes, bear markets will hurt psychologically. But we can lessen the pain significantly if we think of them as an admission fee for worthwhile long-term returns instead of a fine by the market-gods. 

---

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I currently have no vested interest in any company mentioned. Holdings are subject to change at any time.

Last month, I published The Latest Thoughts From American Technology Companies On AI (2024 Q1). In it, I shared commentary in earnings conference calls for the first quarter of 2024, from the leaders of technology companies that I follow or have a vested interest in, on the topic of AI and how the technology could impact their industry and the business world writ large. 

A few more technology companies I’m watching hosted earnings conference calls for 2024’s first quarter after I prepared the article. The leaders of these companies also had insights on AI that I think would be useful to share. This is an ongoing series. For the older commentary:

• 2023 Q1 – here and here
• 2023 Q2 – here and here
• 2023 Q3 – here and here
• 2023 Q4 – here and here

Here they are, in no particular order:

Adobe (NASDAQ: ADBE)

Adobe’s management thinks that creativity is a human trait and that AI assists and amplifies human ingenuity without replacing it

Adobe’s highly differentiated approach to AI is rooted in the belief that creativity is a uniquely human trait and that AI has the power to assist and amplify human ingenuity and enhance productivity.

Adobe’s Firefly generative AI models within its Creative Cloud suite were trained on proprietary data; Adobe’s management has infused AI functionality into its flagship products within the Creative Cloud suite; management has built Adobe Express as an AI-first application; Firefly has generated over 9 billion images since its launch in March 2023 (was 6.5 billion in 2023 Q4); customers are excited about the commercial safety of Firefly; Firefly Services can create significantly more asset variations in a much shorter time, and the speed enables Adobe to monetise Firefly through the volume of content created; Firefly generations in May 2024 was the most generations of any month-to-date; Firefly Services has started to see customer wins; Firefly Services allows users to build customer models and access APIs and these are early in adoption, but customer-interest is better than expected; early Firefly Services usage is on (1) creating multiple variations in the ideation process, (2) creating geography-based variations on ads, (3) assets for community engagement

In Creative Cloud, we’ve invested in training our Firefly family of creative generative AI models with a proprietary data set and delivering AI functionality within our flagship products, including Photoshop, Illustrator, Lightroom, and Premier. We’re reimagining creativity for a broader set of customers by delivering Adobe Express as an AI-first application across the web and mobile surfaces. Since its debut in March 2023, Firefly has been used to generate over 9 billion images across Adobe creative tools…

…This week’s Design Made Easy event, which focused on Express for Business, was another big step forward for us. Companies of all sizes are excited about the integrated power and commercial safety of Firefly, the seamless workflows with Photoshop, Illustrator and Adobe Experience Cloud, and enterprise-grade brand controls that are now part of Express for Business, making it the optimal product for marketing, sales and HR teams to quickly and easily create visual content to share…

… Firefly Services can power the creation of thousands of asset variations in minutes instead of months and at a fraction of the cost. This allows us to monetize the volume of content being created through automation services. The increasing availability of Firefly in Creative Cloud, Express, Firefly Services and the web app is giving us opportunities to access more new users, provide more value to existing users and monetize content automation. These integrations are driving the acceleration of Firefly generations with May seeing the most generations of any month to date…

…On the models, we released Firefly Services. We’ve started to see some customer wins in Firefly Services. So they’re using it for variations, and these are the custom models that we’re creating as well as access to APIs. I would say that’s early in terms of the adoption, but the interest as customers say how they can ingest their data into our models as well as custom models, that’s really ahead of us, and we expect that to continue to grow in Q3 and Q4…

…In terms of what I would say we’re seeing usage of, I think the initial usage of Firefly Services in most companies was all around ideation, how can they create multiple variations of them and in the ideation process really just accelerate that ideation process? Most companies are then starting with as they’re putting it into production, how can they, with the brand assets and the brand guidelines that they have, do this in terms of the variations, whether they be geographic variations or they be just variations. I mean, if you take a step back also, every single ad company right now will tell you that the more variance that you provide, the better your chances are of appropriately getting an uplift for your media spend. So I would say that most companies are starting with creating these variations for geographies. The other one that we see a fair amount of is engaging with their communities. So when they want their communities to have assets that they have blessed for usage within community campaigns, that’s the other place where Firefly Services are being used. And a company has a community portal where the community can come in, take something and then post whether it’s on whatever social media site that you want. 

Adobe’s management has introduced Acrobat AI Assistant, an AI-powered tool for users to have conversations with their documents, within Adobe’s Document Cloud suite; Acrobat AI Assistant features are available as a standalone offer or as an add-on subscription to existing Adobe products; Acrobat AI Assistant for English documents was made generally available in April; management is seeing early success in adoption of Acrobat AI Assistant; Acrobat AI Assistant can be applied to document types beyond PDFs

In Document Cloud, we’re revolutionizing document productivity with Acrobat AI Assistant, an AI-powered conversational engine that can easily be deployed in minutes. This enhances the value of the trillions of PDFs, which hold a significant portion of the world’s information. Acrobat AI Assistant features are now available through an add-on subscription to all Reader and Acrobat enterprise and individual customers across desktop, web and mobile…

…The introduction of Acrobat AI Assistant made generally available in April for English documents marks the beginning of a new era of innovation and efficiency for the approximately 3 trillion PDFs in the world. Acrobat AI Assistant is empowering everyone to shift from reading documents to having conversations with them in order to summarize documents, extract insights, compose presentations and share learnings. AI Assistant is available as a stand-alone offer for use in reader and as an add-on to Acrobat Standard and Pro. We’re seeing early success driving adoption of AI Assistant as part of our commerce flows and remain optimistic about the long-term opportunities…

…Other business highlights include general availability of Acrobat AI Assistant support for document types beyond PDF, meeting transcripts and enterprise requirements. 

The Adobe Experience platform, which is part of the Digital Experience segment, is on track to become a billion-dollar annual revenue business; management has released AEP (Adobe Experience Platform) AI Assistant to improve the productivity of marketing professionals; Adobe is the #1 digital experience platform; customer interest and adoption of AEP AI Assistant is great

At the end of May, we celebrated the 5-year anniversary of Adobe Experience Platform, which we conceived and built from scratch and which is on track to be the next billion-dollar business in our Digital Experience portfolio. We released AEP AI Assistant to enhance the productivity of marketing practitioners through generative AI while expanding access to native AEP applications…

…When we introduced Adobe Experience Platform 5 years ago, it was a revolutionary approach to address customer data and journeys. Today, we’re the #1 digital experience platform and AEP with native apps is well on its way to becoming a billion-dollar business…

…We are excited by the customer interest and adoption of our latest innovations, including AEP AI Assistant, a generative AI-powered conversational interface that empowers practitioners to automate tasks, simulate outcomes and generate new audiences and journeys. For example, customers like General Motors and Hanesbrands have been working with AEP AI Assistant to boost productivity and accelerate time to value while democratizing access to AEP and apps across their organizations…

…When you think about the AEP AI Assistant, it’s doing a couple of things. One, it’s really making it easier for customers to deploy use cases. When you think of use cases that they have around, for example, generating audiences and running campaigns around those audiences, these are things today that require some data engineering. They require the ability to put these audiences together. So they require marketing and IT teams to work together. The AEP AI Assistant is making it much easier for marketers to be able to do it themselves and be able to deploy a lot more use cases.

Adobe’s management’s vision for Adobe Express is to make design easy; the launch of the new Adobe Express app in 2024 Q1 (FY2024 Q2) has been well received, with monthly active users doubling sequentially; management has been deeply integrating AI features into Adobe Express; cumulative exports from Adobe Express has increased by 80% year-on-year in 2024 Q1; management is building Adobe Express to be AI-first; management thinks Adobe Express is leveraging people’s need for AI 

Our vision for Adobe Express is to provide a breakthrough application to make design easy for communicators worldwide, leveraging generative AI and decades of Adobe technology across web and mobile. Our launch of the all-new Express application on iOS and Android earlier this quarter is off to a strong start with monthly active users doubling quarter-over-quarter…

…There’s a lot of buzz with Express here at Adobe coming off the event we just had earlier this week, but it’s really based on the fact that the innovation in Express is on a tear, right? A few months ago, we introduced an all-new Express for the web. This quarter, we introduced an all-new Express for mobile. We introduced Express for Business. We also now have, as we’ve just talked about, been more deeply integrating AI features, whether it’s for imaging generation or Generative Fill or text effects, character animation, design generations, more deeply into the flow for Express. And that combination has led to an incredible set of metrics over the last quarter, in particular, but building throughout the year. Express MAU is growing very quickly. We talked about in the script earlier that MAU on mobile has more than doubled quarter-over-quarter, which is fantastic to see. And cumulative exports, if you look at year-over-year, has grown by over 80%. So really feeling good about sort of the momentum we’re seeing…

…Express that is now in market is built on a brand-new platform, right? And that brand-new platform lays the groundwork for the AI era. And this will be — Express will be the place that anyone can come and create through a combination of conversational and standard inputs. That’s the vision that we have. And I think it’s an opportunity for us to really leap forward in terms of what we can do on the web and mobile at Adobe…

…Express is really being driven by sort of the need for AI and how people are able to describe what they want and get the final output. When David talked about exports, just to clarify, what that means is people who have successfully got what they want to get done, done. And that’s a key measure of how we are doing it, and AI is certainly facilitating and accelerating that.

Adobe GenStudio uses AI to help enterprises transform their content supply chain; enterprise customers view customer experience management and personalisation at scale as key investments to make

We’re now transforming the content supply chain for enterprises with Adobe GenStudio, enabling them to produce content at scale, leveraging generative AI through native integrations with Firefly Services and Adobe Express for Business. Enterprise customers, both B2C and B2B, view customer experience management and personalization at scale as key areas of differentiation, making it a priority investment for Chief Marketing Officers, Chief Information Officers and Chief Digital Officers.

Adobe’s management thinks the biggest opportunity in AI for Adobe is in interfaces, such as performing tasks faster, improving workflows etc; in AI interfaces, management is seeing significant usage in AI Assistant and Photoshop; management believes that (1) the real benefits from disruptive technologies such as AI come when people use interfaces to improve their work, and that (2) in the future, more people will be using these interfaces

I think the biggest opportunity for us and why we’re really excited about GenAI is in the interfaces because that’s the way people derive value, whether it’s in being able to complete their tasks faster, whether it’s be able to do new workflows. And I would say, in that particular space, Acrobat has really seen a significant amount of usage as it relates to AI Assistant and Photoshop…

… And so we’re always convinced that when you have this kind of disruptive technology, the real benefits come when people use interfaces to do whatever task they want to do quicker, faster and when it’s embedded into the workflows that they’re accustomed to because then there isn’t an inertia associated with using it…

…And so net-net, I am absolutely betting on the fact that 5 years from now, there’ll be more people saying, “I’m using creative tools to accomplish what I want,” and there’ll be more marketers saying, “I can now, with the agility that I need, truly deliver a marketing campaign in an audience that’s incredibly more specific than I could in the past.” And that’s Adobe’s job to demonstrate how we are both leading in both those categories and to continue to innovate.

Adobe’s management’s primary focus for generative AI is still on user adoption and proliferation

From the very beginning, we’ve talked to you guys about our primary focus for generative AI is about user adoption and proliferation, right? And that has continued to be the primary thing on our mind.

Adobe’s management thinks there are different routes to monetise AI, such as winning new users, and getting higher ARPU (average revenue per user)

And to your point, there are many different ways that we can monetize this. First is as you think about the growth algorithms that we always have in our head, it always starts with, as Shantanu said, new users, right? And then it’s about getting more value to existing users at higher ARPU, right? So in the context of new users, first and foremost, we want to make sure that everything we’re doing generative AI is embedded in our tools, starting with Express, right?

Adobe has seen strong growth in emerging markets because users need access to the cloud for all of the AI functionality

I mean I think in the prepared remarks, Dan also talked about the strength in emerging markets. And I think the beautiful part about AI is that since they need access to the cloud to get all of the AI functionality, emerging market growth has been really strong for us.

Adobe’s management thinks that they have hit a sweet spot with pricing for generative AI credits in Adobe’s subscription plans for imaging and vector work, but they will need to explore different plans for generative AI credits when it comes to video work

When we think about what we’ve done with imaging and video, we’ve done the right thing by making sure the higher-value paid plans that people don’t have to think about the amount of generative capability. And so there, the balance between for free and trialist users, they’re going to run into the generative capability limits and therefore, have to subscribe. But for the people who actually have imaging and vector needs, that they’re not constantly thinking about generative, I think we actually got it right. To your point, as we move to video, expect to see different plans because those plans will, by necessity, take into account the amount of work that’s required to do video generation. So you’re absolutely right as a sort of framework for you to think about it.

Adobe’s management thinks that there’s a lot of excitement now on AI infrastructure and chips, but the value of AI will need to turn to inference in order for all the investment in AI infrastructure and chips to make sense

It’s fair to say that the interest that exists right now from investors, as it relates to AI, is all associated with the infrastructure and chips and perhaps rightly so because that’s where everybody is creating these models. They’re all trying to train them. And there’s a lot of, I think, deserved excitement associated with that part of where we are in the evolution of generative AI. If the value of AI doesn’t turn to inference and how people are going to use it, then I would say all of that investment would not really reap the benefit in terms of where people are spending the money.

Adobe’s management think it doesn’t matter what AI model is used to generate content – DALL-E, Firefly, Midjourney, or more – because the content ultimately needs to be edited on Adobe’s software; management is building products on Firefly, but they are also happy to leverage on third-party AI models

So Firefly might be better at something. Midjourney might be something at something else. DALL·E might do something else. And the key thing here is that, around this table, we get excited when models innovate. We get excited when Firefly does something amazing. We get excited when third-party models do something because our view, to Shantanu’s point, is that the more content that gets generated out of these models, the more content that needs to be edited, whether it’s color correction, tone matching, transitions, assembling clips or masking compositing images. And the reason for this is that this is not a game where there’s going to be one model. There’s — each model is going to have its own personality, what it generates, what it looks like, how fast it generates, how much it costs when it generates that, and to have some interface layer to help synthesize all of this is important. And so just sort of to note, we’ve said this before but I’ll say it again here, you will see us building our products and tools and services leveraging Firefly for sure, but you’ll also see us leveraging best-of-breed personalities from different models and integrate them all together.

Ultimately, generative AI is going to create more growth in Adobe’s category

[Analyst] Awesome, the message here is that GenAI is going to create more growth in the category. And Shantanu, you did that with the pivot to cloud. You grew the category, so here we go again.

DocuSign (NASDAQ: DOCU)

DocuSign Navigator is a new AI-powered product that allows users to store and manage their entire library of accumulated agreements, including non-DocuSign agreements

Second, DocuSign Navigator allows you to store, manage and analyze the customer’s entire library of accumulated agreements. This includes past agreements signed using DocuSign eSignature as well as non-DocuSign agreements. Navigator leverages AI to transform unstructured agreements into structured data, making it easy to find agreements, quickly access vital information, and gain valuable insights from agreements. 

DocuSign acquired Lexion, an AI-based agreements company, this May; management thinks Lexion can improve Docusign’s Agreement AI and legal workflow; the Lexion acquisition is not for revenue growth, but to integrate the AI technology into DocuSign’s products

AI is central to our platform vision, and we’re thrilled to welcome Lexion to the DocuSign family. Lexion is a proven leader in AI-based agreement technology, which significantly accelerates our IAM platform goals. We maintain a high bar for acquisitions, and Lexion stood out due to its sophisticated AI capabilities, compatible technology architecture, and promising commercial traction with excellent customer feedback, particularly in the legal community…

… With regard to capital allocation, we also closed the Lexion acquisition on May 31…

…In terms of how it adds to DocuSign, I think overall, agreement AI, their extraction quantity and quality where we augment our platform. Another area where I think they’re really market-leading is in legal workflow. So workflow automation for lawyers, for example, if you’re ingesting a third-party agreement, how can you immediately use AI to assess the agreement, understand how terms may deviate from your standard templates and highlight language that you might want to propose as a counter that really accelerates productivity for legal teams. And they’ve done an excellent job with that. So overall, that’s how it fits in…

…We’re not breaking it out just because of its size and materiality. It’s not material to revenue or op margin for us. The overarching message that I would like to send on Lexion is that the purchase of Lexion is about integrating the technology into the DocuSign IAM platform. That opportunity for us, we think, in the long term, can apply to the well over 1 million customers that we have.

MongoDB (NASDAQ: MDB)

MongoDB’s management wants to prioritise investments in using generative AI to modernise legacy relational applications; management has found that generative AI can help with analyzing existing code, converting existing code and building unit and functional tests, resulting in a 50% reduction in effort for app modernisation; management sees a growing list of customers across industries and geographies who want to participate; interest in modernising legacy relational applications is high, but it’s still early days for MongoDB

Second, we are more optimistic about the [ opti-tech ] to accelerate legacy app modernization using AI. This is a large segment of the market that has historically been hard to penetrate. We recently completed the first 2 GenAI powered modernization pilots, demonstrating we can use AI to meaningfully reduce the time, cost and risk of modernizing legacy relational applications. In particular, we see that AI can significantly help with analyzing existing code, converting existing code and building unit and functional tests. Based on our results from our early pilots, we believe that we may be able to reduce the effort needed for app modernization by approximately 50%. We have a growing list of customers across different industries and geos, who want to participate in this program. Consequently, we will be increasing our level of investment in this area…

…We have an existing relational migrated product that allows people to essentially migrate data from legacy relational databases and does the schema mapping for them. The one thing it does not do, which is the most cumbersome and tedious part of the migration is to auto generate or build application code. So when you go from a relational app to an app built on MongoDB, you still have to essentially rewrite the application code. And for many customers, that was the inhibitor for them to migrate more apps because that takes a lot of time and a lot of labor resources. So our app modernization effort is all about or using AI is all about now solving the third leg of that stool, which is being able to reduce the time and cost and effort of rewriting the app code, all the way from analyzing existing code, converting that code to new code and then also building the test suites, both unit tests and functional tests to be able to make sure the new app is obviously operating and functioning the way it should be…

…That’s why customers are getting more excited because the lower you reduce the cost for that migration or the switching costs, the more apps you can then, by definition, migrate. And so that is something that we are very excited about. I will caution you that it’s early days. You should not expect some inflection in the business because of this. 

MongoDB’s management wants to prioritise investments in building an ecosystem for customers to build AI-powered applications because management recognises that there are other critical elements in the AI tech stack beyond MongoDB’s document-based database; management has launched the MongoDB AI Application Program, or MAP, that combines cloud computing providers, model providers, and more; Accenture is the first global systems integrator to join MAP

Third, although still early in terms of customers building production-ready AI apps, we want to capitalize on our inherent technical advantages to become a key component of the emerging AI tech stack…

… Recognizing there are other critical elements of the AI tech stack, we are leveraging partners to build an ecosystem that will make it easier for customers to build AI-powered applications. Earlier this month, we launched the MongoDB AI application Program, or MAP, a first-of-its-kind collaboration that brings together all 3 hyperscalers, foundation model providers, generative AI frameworks, orchestration tools and industry-leading consultancies. With MAP, MongoDB offers customers reference architectures for different AI use cases, prebuilt integrations and expert professional services to help customers get started quickly. Today, we are announcing that Accenture is the first global systems integrator to join MAP and that it will establish a center of excellence focused on MongoDB projects. We will continue to expand the program through additional partnerships and deeper technical integrations.

MongoDB’s document-based database architecture is a meaningful differentiator in AI because AI use cases involve various types of data, which are incompatible with legacy databases; there was a customer who told management that if he were to design a database specifically for AI purposes, it would be exactly like MongoDB

Customers tell us that our document-based architecture is a powerful differentiator in an AI world, the most powerful use cases rely on data of different types and structures such as text, image, audio and video. The flexibility required to handle a variety of different data structures is fundamentally at odds with legacy databases that rely on rigid schemes, which is what makes MongoDB’s document model such a good fit for these AI workloads…

…One customer told us if he had to build a database, it would be designed exactly like MongoDB and so for this new AI era. And so we feel really good about our position. 

A unit with Toyota that is focused on AI and data science migrated to MongoDB Atlas after experiencing reliability issues with its original legacy database system; the Toyota unit now uses MongoDB Atlas for over 150 micro-services and will use MongoDB Atlas as its database of choice for future AI needs

Toyota Connected, an independent Toyota company focused on innovation, AI, data science, and connected intelligence services, migrated to MongoDB Atlas after experiencing reliability issues with the original legacy database system. The team selected MongoDB Atlas for its ease of deployment, reliability and multi-cloud and multi-region capabilities. Toyota Connected now uses Atlas for over 150 micro-services. Their solution benefits from 99.99% uptime with Atlas as a platform for all data, including mission-critical vehicle telematics and location data needed for emergency response services. MongoDB’s Toyota Connected’s database of choice for all future services as they explore vector and AI capabilities, knowing they’ll get the reliability and scalability they need to meet customer needs.

Novo Nordisk is using MongoDB Atlas Vector Search to power its generative AI efforts in producing drug development reports; Novo Nordisk switched from its original relational database when it wasn’t capable of handling complex data and lacked flexibility to keep up with rapid feature development; reports that Novo Nordisk used to take 12 weeks to prepare can now be completed with MongoDB Atlas Vector Search in 10 minutes

By harnessing GenAI with MongoDB Atlas Vector search, Novo Nordisk, one of the world’s leading health care companies is dramatically accelerating how quickly can get new medicines approved and delivered to patients. The team responsible for producing clinical study report turn to Atlas when the original relational database wasn’t capable of handling complex data and lack the flexibility needed to keep up with the rapid feature development. Now with GenAI and the MongoDB Atlas platform, Novo Nordisk gets the mission-critical assurances that needs to run highly regulated applications, enabling them to generate complete reports in 10 minutes rather than 12 weeks. 

MongoDB’s management still sees MongoDB as well-positioned to be a key beneficiary when organisations embed AI into next-gen software applications

Our customers recognize that modernizing legacy applications is no longer optional in the age of AI. And are preparing for a multiyear journey to accomplish that goal. They see MongoDB as a key partner in that journey. We are well positioned to be a key beneficiary as organizations embed AI into the next generation of software applications that transform their business.

MongoDB’s management  believes that MongoDB’s performance in 2024 Q1 was less upbeat than the cloud computing hyperscalers because the hyperscalers’ growth came primarily from reselling GPU (graphic processing unit) capacity for AI training and there’s a lot of demand for AI training at the moment, whereas MongoDB is not seeing AI apps in production at scale, which is where MongoDB is exposed to

In contrast to the hyperscalers, like we believe the bulk of their growth across all 3 hyperscalers was really spent on reselling GPU capacity because there’s a lot of demand for training models. We don’t see a lot of, at least today, a lot of AI apps in production. We see a lot of experimentation, but we’re not seeing AI apps in production at scale. And so I think that’s the delta between the results that the hyperscalers produce versus what we are seeing in our business.

MongoDB’s management  thinks that AI is going to drive a step-fold increase in the number of apps and software created, but it’s going to take time, although the process is happening

I think with AI, you’re going to see a stepfold increase in the number of apps and the number of amount of software that’s being built to run businesses, et cetera. But that’s going to take some time. as with any new adoption cycle, the adoption happens in what people commonly refer to as S curves. And I think we’re going through one of those S curves.

MongoDB’s management sees the possibility of customers’ desire to spend on AI crowding out other software spending, but does not think it is an excuse for MongoDB not meeting new business targets

Is AI essentially crowding out new business? We definitely think that that’s plausible. We definitely see development teams experimenting on AI projects. The technology is changing very, very quickly. But that being said, we don’t see that as a reason for us to not hit our new business targets. And as I said, even though we started slow, we almost caught up at the end of this quarter, and we feel really good about our new business opportunity for the rest of this year. So — so I don’t want to use that as an excuse for us not meeting our new business targets.

Okta (NASDAQ: OKTA)

A new product, Identity Threat Protection with Okta AI, is expected to become generally available soon

We’re also excited about the launch of Identity Threat Protection with Okta AI, which includes powerful features like Universal Logout, which makes it possible to automatically log users out of all of their critical apps when there is a security issue. Think of this as identity threat detection and response for Okta. We expect Identity Threat Protection to become generally available this summer.

Okta’s management does not expect the company’s new products – which includes governance, PAM, Identity Threat Protection with Okta AI, and Identity Security Posture Management – to have material impacts on the company’s financials in FY2025; of the new products, management thinks Identity Threat Protection with Okta AI and Identity Security Posture Management will make impacts first before PAM does

I wouldn’t expect for these newer things that are coming out like posture management or threat protection, I wouldn’t expect it in FY ’25 at all. I probably wouldn’t even think it would impact it in FY ’26 because we’re talking about a $2.5 billion business at this point. It takes a lot of money in any of these products to make a material difference to the overall numbers. So we’re setting these up for the long term…

…How we’re thinking about this internally is that the — I think it will mirror the order of broad enablement. So we’re broadly enabling people in the following order: governance is first, followed by a combination of posture management and identity threat protection, followed by privileged access. So we think that Identity Threat Protection with Okta AI and Identity Security Posture Management, that bundle could pretty quickly have as much of an impact as governance. And then we think the next sequential enablement in the next order of impact will probably be Privileged Access.

Okta’s management is currently not seeing companies changing their software spending plans because they want to invest in AI, although that might change in the future

[Question] There is a shift in the marketplace among the C-suite from fear about the economy to, gee, I need to focus on how I’m going to implement AI. And in that context, there’s uncertainty around the mechanics of what they need to do to secure AI within their organizations. And I guess my question to you is we’re hearing the pipelines of the VAR channels, particularly in security, are extremely robust into the back half of the year. But the uncertainty around AI decision is keeping people from implementing it. So how robust is the pipeline that you’re looking at? And are you, in fact, hearing that from your C-suite customers when you talk to them?

[Answer] What I’ve heard is everyone is figuring out how they can deploy this new wave of technology to their products and services and business and how they can use it for security and how they can use it for innovation. But they’re not at the stage where it’s broadly impacting other plans. It’s more of like a — their planning exercise at this point. I think that might change in the future.

Okta’s management thinks that more companies will invest in AI in the future, and this will be a tailwind for the company because more identity features will be needed; the current AI wave is not impacting spending on Okta at the moment, but might be a boon in the future

My bet is that they’re going to be building new apps. They’re going to be deploying more technology from vendors that are building apps with AI built in, which is going to — all of that’s going to lead to more identity. They’re going to have to log people into their new apps they build. They’re going to have to secure the privileged accounts that are running the infrastructure behind the new apps. They’re going to have to make sure that people in their workforce can get to the apps that are the latest, greatest AI-driven experiences for support or for other parts of the business. So I think that identity is one of these foundational things that’s going to be required whether it’s the AI wave, which is going to be really real and impactful and — or whether it’s whatever comes after that.

[Question] So not impacting spending today but might impact to help it in the future.

[Answer] Yes, yes. That’s how I see it.

Okta’s management sees 2 ways of monetising Okta AI: Through new products, and through making existing products better

Okta AI will be monetized through 2 ways. one will be new products like Identity Threat Protection with Okta AI; and the other way, it will be — it will just make products better. For example, the Identity Security Posture Management, it has a new capability that’s going to be added to that product that’s just going to make it smarter about how it detects service accounts. That Identity Security Posture Management scans a customer’s entire SaaS estate, and says, here are all the things you should look at. You should take — this account needs MFA. This other account is — probably has overly permissive permissions. The challenge there is how does the customer know which of those accounts are service accounts, so they can’t have human biometrics. And we added — we used some AI capability to add that to the scan. So that’s an example of just the product gets better versus Identity Threat Protection is like it’s a whole new product enabled by that.

Salesforce (NYSE: CRM)

Salesforce is managing 250 petabytes of customer data and management thinks this is going to be critical and positions Salesforce for success when Salesforce’s customers move into AI; management thinks that customer data is the critical success factor in AI, not AI models and UIs (user interfaces); management thinks most of the AI models that are being built today, both large and small, are just commodities and will not survive

We’re now managing more than 250 petabytes of data for our customers. This is going to be absolutely critical as they move into artificial intelligence…

…When you look at the power of AI, you realize the models and the UI are not the critical success factors. It’s not critical where the enterprise will transform. There are thousands of these models, some open source and some close source models, some built with billions, some with just a few dollars, most of these will not survive. They’re just commodities now, and it’s not where the intelligence lies. And they don’t know anything about a company’s customer relationships. Each day, hundreds of petabytes of data are created that AI models can use for training and generating output. But the one thing that every enterprise needs to make AI work is their customer data as well as the metadata that describes the data, which provides the attributes and contacts the AI models need to generate accurate, relevant output. And customer data and metadata are the new gold for these enterprises…

…Not every company is as well positioned, as you know, for this artificial intelligence capability of Salesforce is because they just don’t have the data. They may say they have this capability or that capability, this user interface, that model, that whatever, all of these things are quite fungible and are expiring quickly as the technology rapidly moves forward. But the piece that will not expire is the data. The data is the permanent key aspect that, as we’ve said, even in our core marketing, it’s the gold for our customers and their ability to deliver our next capability in their own enterprises.

Salesforce’s management is seeing incredible momentum in Data Cloud, which is Salesforce’s fastest-growing organic and next billion-dollar cloud; Data Cloud’s momentum is powered by the need for customers to free their data from being trapped in thousands of apps and silos; the need to free their data is important if Salesforce’s customers want to embrace AI; Data Cloud was in 25% of Salesforce’s >$1 million deals in 2024 Q1; 2024 Q1 was the second quarter in a row when >1,000 Data Cloud customers were added; in 2024 Q1, 8 trillion records were ingested in Data Cloud, up 42% year-on-year, 2 quadrillion records were processed, up 217% year-on-year, and there were 1 trillion activations, up 33% year-on-year

Many of these customers have a central business and customer data that exists outside of Salesforce that’s trapped in thousands of apps and silos. It’s disconnected. That’s why we’re seeing this incredible momentum with our Data Cloud, our fastest-growing organic, and our next billion-dollar cloud. It’s the first step to becoming an AI enterprise. Data Cloud gives every company a single source of truth and you can securely power AI insights and actions across the entire Customer 360.

Now let me tell you why I’m excited about Data Cloud and why it’s transforming our customers and how it’s preparing them for this next generation of artificial intelligence. Data Cloud was included in 25% of our $1 million-plus deals in the quarter. We added more than 1,000 Data Cloud customers for the second quarter in a row. 8 trillion records were ingested in the Data Cloud in the quarter, up 42% year-over-year, and we processed 2 quadrillion records. That’s a 217% increase compared to last year. Over 1 trillion activations drove customer engagement, which is a 33% increase year-over-year. This incredible growth of data in our system and the level of transactions that we’re able to deliver not just in the core system but especially in data cloud is preparing our customers for this next generation of AI.

Salesforce’s predictive AI, Einstein, is generating hundreds of billions of predictions daily; Salesforce is working with thousands of customers in generative AI use cases through the launch in 2024 Q1 of Einstein Copilot, Prompt Builder,and Einstein Studio; Salesforce has closed hundreds of Einstein Copilot deals since the product’s general availability (GA)

Einstein is generating hundreds of billions of predictions per day, trillions per week. Now we’re working with thousands of customers to power generative AI use cases with our Einstein Copilot, our Prompt Builder, our Einstein Studio, all of which went live in the first quarter, and we’ve closed hundreds of Copilot deals since this incredible technology has gone GA. And in just the last few months, we’re seeing Einstein Copilot develop higher levels of capability. We are absolutely delighted and could not be more excited about the success that we’re seeing with our customers with this great new capability.

Luxury fashion company Saks is using Salesforce’s Einstein 1 Platform in Data Cloud to create AI-powered personal experiences for customers

Saks, a leader in the luxury fashion market, part of Hudson’s Bay, went all-in on Salesforce in the quarter. CEO Marc Metrick is using AI to create more personal experiences for every customer touch point across their company. And with our Einstein 1 Platform in Data Cloud, Saks can unify and activate all its customer data to power trusted AI.

Salesforce is helping FedEx generate savings and accelerate top-line partly with the help of its AI solutions

The Salesforce data and app and AI capabilities generate expense savings. This is the core efficiency while growing and accelerating top line revenue. This is the effectiveness that we’re delivering for FedEx. This efficiency includes next best action for sellers, automated lead nurturing, Slack for workflow management, opportunity scoring, a virtual assistant, AI on unstructured data for delivering content to sales and customer service. And when we think about effectiveness, we see our Journey Builder delivering hyper personalization, integrating customer experiences across service, sales, marketing, the ability to tailor and deliver customer experiences based on a Customer 360 view. When we look at these incredible next generation of capability we’ve delivered at FedEx, gone now are these days of static business rules that leave customers dissatisfied, asking, “Do they not know that I’m a valued customer of FedEx?” Now FedEx has not only the power of the Customer 360 but the power of AI to unlock so much more commercial potential by conducting an orchestra of commercial functions that never played well together before.

Air India is using Data Cloud and Einstein across 550,000 service cases each month to improve its customer experience and deliver more personalised customer service

And with Data Cloud, Air India is unifying Data Cloud across loyalty, reservations, flight systems and data warehouses. They have a single source of truth to handle more than 550,000 service cases each month. And now with Einstein, we’re automatically classifying and summarizing cases and sending that to the right agent who’d recommend the next steps and upgrading in high-value passenger experiences. Even when things happen like a flight delay, our system is able to immediately intervene and provide the right capability to the right customer at the right time. All of that frees up agents to deliver more personal service and create more personal relationships, a more profitable, a more productive, a more efficient Air India, a company that’s using AI to completely transform their capability.

Salesforce’s management is seeing good demand, driven by customers recognising the value of transforming their front-office operations with AI, but buying behaviour among customers is measured (similar to the past 2 years) with the exception of 2023 Q4

We’re seeing good demand as AI technology rapidly evolves and customers recognize the value of transforming into AI enterprises. CEOs and CIOs are excited about the opportunity with data and AI and how it can impact their front-office operations…

…We continue to see the measured buying behavior similar to what we experienced over the past 2 years and with the exception of Q4 where we saw stronger bookings. The momentum we saw in Q4 moderated in Q1 and we saw elongated deal cycles, deal compression and high levels of budget scrutiny.

Siemens used Einstein 1 Commerce to build and launch its AI-powered digital marketplace, named Accelerator Marketplace, in just 6 months

Siemens lacked a centralized destination for customers to easily choose the right products and buy on demand. To simplify the buying experience for customers, Siemens worked with Salesforce to develop and launch its Accelerator Marketplace, an AI-powered digital marketplace built on Einstein 1 Commerce, providing AI-generated product pages, smart recommendations and self-service ordering. And they did it all in just 6 months.

Salesforce is using AI internally with great results; Salesforce has integrated Einstein into Slack and Einstein has already answered 370,000 employee queries in a single quarter; Salesforce’s developers have saved 20,000 hours of coding through the use of AI tools

AI is not just for our customers. As part of our own transformation, we continue to adopt AI inside Salesforce. Under the leadership of our Chief People Officer Nathalie Scardino and our Chief Information Officer Juan Perez, we’ve integrated Einstein right into Slack, helping our employees schedule, plan and summarize meetings and answer employee questions. Einstein has already answered nearly 370,000 employee queries in a single quarter. In our engineering organization, our developers now save more than 20,000 hours of coding each month through the use of our AI tools.

Slack AI was launched in February and it provides recap, summaries and personalized search within Slack; >28 million Slack messages have been summarised by Salesforce’s customers since the launch of Slack AI

We also launched Slack AI in February, an amazing innovation that provides recap, summaries and personalized search right within Slack. I personally have been using it every day to get caught up on the conversations happening in every channel. And we’ve seen great traction with our customers with this product, and our customers have summarized over 28 million Slack messages since its launch in February.

Los Angeles city will use Salesforce’s Government Cloud and other solutions to integrate AI into its software system

And in the public sector, the city of Los Angeles chose Salesforce to modernize how the city’s 4 million residents request city services using its MyLA311 system. The city will use government cloud and other Salesforce solutions to integrate AI assistance into MyLA311 and modernize its own constituent-facing services, giving residents more self-service options and improving service reliability and responsiveness.

Salesforce’s products for SMBs (small and medium businesses), Start and Pro Suite, which both have AI built-in, are building momentum; Salesforce added 2,300 new logos to the products in 2024 Q1

Our new offerings for small and medium businesses, Starter and Pro Suite, which are ready-to-use, simplified solutions, with AI built in, are building momentum. In Q1, we added another 2,300 new logos to these products. Since Starter’s launch last year, we’ve seen customers upgrade to our recently launched Pro Suite and even to our Enterprise and Unlimited editions.

Studies have shown that 75% of the value of generative AI use cases is in the front office of companies; Salesforce is the leader in front-office software, so management thinks this is why – with Data Cloud at the heart – the company is in a good position for growth going forward

We all saw the report from McKinsey, 75% of the value of Gen AI use cases is in the front office. And everybody knows Salesforce is the leader in front-office software. That’s our fundamental premise for our growth going forward. We’re already managing 250 petabytes of data and metadata that’s going to be used to generate this incredible level of intelligence and artificial intelligence capability to deliver for our customers a level of productivity and profitability they’ve just never been able to see before. And at the heart of that is going to be our Data Cloud. 

Salesforce’s management is focused on 2 things at the company: The ongoing financial transformation at Salesforce, and the use of AI

Look, we really are focused on two things in our company. One is this incredible financial transformation that we’ve all gone through with you in the last year. The second one is this incredible transformation to artificial intelligence, which is going to be based on data. 

Salesforce’s management thinks that the relative weakness seen in the software world currently is because of pull-forward in demand from COVID, and not because of crowding out by AI; management thinks AI is a growth driver for software companies

[Question] When we think about this measured buying environment, is there any sort of crowding effect around AI that’s impacting software in your view, meaning when you think about all these companies starting to gear up for this next platform shift, was it just the uncertainty of what they’re going to spend on over the next 6 to 12 months, holding them back perhaps on what their normal sort of pace of spending might be with you all or other enterprise software companies?

[Answer] As we entered the post-pandemic reality, we saw companies who had acquired so much software in that time looked to actually rationalize it, ingest it, integrate it, install it, update it. I mean it’s just a massive amount of software that was put in. And so every enterprise software company kind of has adjusted during end of this post-pandemic environment. So when you look at all of these companies, especially as you saw them report in the last 30 days, they’re all basically saying that same thing in different ways. When you take AI, that has to be our growth driver for future capabilities for these companies. 

Salesforce’s management sees the consumer AI world and the enterprise AI world as having very different needs for the kind of data they use for AI implementations, even though the model architectures are very similar; enterprise AI requires internal datasets from companies

It’s been pretty magical to use OpenAI over the last year, especially in the last release, when I’m really talking to it. And when I think about the incredible engineering effort that OpenAI has done, it’s pretty awesome. They’ve built a great UI. I love talking to the software. They have really strong algorithms or what we call Models, especially their new one, which is their 4o Model. And then they stole data from lots of companies like Time, Dow Jones, New York Times, Reddit. Now they’re all making good, doing agreements with all of us, saying, “We’re sorry,” and paying for it. And they took that data, they normalized it, they delivered a comprehensive data set that they train their model on…

…And then we’ve seen a lot of fast followers with the models. It could be open source models like Llama 3. It could be some proprietary models like Gemini from Google and others. Now there’s thousands and thousands of these models. And if you look on Hugging Face, everybody is a fast follower. And 6 months later, everybody is where everybody else was 6 months ago. And the data, well, a lot of these companies are all thinking they can rip off all this data, too, and they’re all having to pay that price. Okay, that’s the consumer world.

The enterprise world is a little different, right? We have great user interfaces, great apps, all kinds of great technology that our users are using, the millions and millions of users. Then we have the same models, in many cases, or maybe we’ve written some of our own models with our engineers. But then the third piece is the data. And that data is a little bit different. Because in the enterprise, how do you put together these large, fully normalized data sets to deliver this incredible capability, and that is where the magic is going to be. Because for all companies, including ours and others, who want to deploy generative AI internally, it’s not going to be Times Magazine that’s going to give you the intelligence, it’s going to be our customer data and your transaction history and how you’re how your company operates in your workflow and your metadata. And that idea that we can deliver another level of productivity for companies using that architecture is absolutely in front of us. But that idea that we have to do it with the right architecture, that also is in front of us. And I think that while we can say it’s a different kind of architecture, it’s still the same idea that we need a great UI, we need models, but we’re going to need very highly normalized and federated data. And that data needs to be stored somewhere, and it needs to come from somewhere. And that is going to be something that’s going to continue in perpetuity over time as these models and UIs are quite fungible. And we’ll be using different models and different UIs over the years, but we’ll be using the same deep data sources. And I think that is why, when I look at what Salesforce is doing, this is going to be critical for our customers.

Salesforce’s management has seen many instances where software vendors promise customers they can deliver AI magic, only for the customers to come up empty-handed because (1) the vendors did not put in the work – and are unable – to make the customers’ data AI-ready, and (2) there’s no proper UI that’s commonly accessed within the customer

Don’t think that there aren’t a lot of people walking into these companies saying, “Hey, you can do this. You can do that. You can do these other things”. We’ve seen a lot of that in the last 6 to 12 months, and then it turns out that you can’t. “Hey, I can make this happen. I can make that happen. I can pull a rabbit out of the hat in the enterprise for you by doing this, that and the other thing,” and then it doesn’t actually happen. And then what it turns out is you got to do a lot of the hard work to make this AI happen, and that starts with building highly normalized, large-scale, federated, highly available data sources. And then building on top of that the kind of capabilities to deliver it to our customers. I think a common story is, “Hey, oh, yes, I am a provider of a data lake or a data capability. And just by going to that, I’m going to be able to provide all your AI.” But then it turns out that no one in the enterprise actually uses that product. There is no UI that’s commonly accessed. That’s why I’m so excited that Salesforce has Sales Cloud and Service Cloud and Tableau and Slack and all of our amazing products that have these huge numbers of users that use these products every single day in a trusted, scalable way and then connecting that into this new capability.

Veeva Systems (NYSE: VEEV)

Veeva’s management’s strategy with generative AI is to enable customers and partners to develop generative AI solutions that work well with Veeva’s applications; generative AI applications require access to data and Veeva provides the access through solutions such as Direct Data API; Direct Data APi provides data access 100 times faster than traditional APIs; management is seeing customers being appreciate of Veeva’s efforts to allow generative AI applications to work well with its own applications; management thinks that the generative AI applications its customers and partners will develop will be very specific; Veeva’s work on Direct Data API started more than 2 years ago

In these early days as GenAI matures, our strategy is to enable our customers and partners to develop GenAI solutions that work well with Veeva applications through our AI Partner Program and powerful Vault Platform capabilities like the Vault Direct Data API. GenAI applications need access to accurate, secure, and timely data from Vault and our data applications. Released in April, our Direct Data API provides data access up to 100 times faster than traditional APIs…

…In general, customers are appreciative of our strategy to enable a broad range of GenAI use cases and experimentation through their own resources and our partner network…

…In terms of the AI strategy, our strategy is to really enable customers and their partners to develop AI applications because they’re going to be very specific AI applications, GenAI applications for very specific use cases whether it’s field information, pre-call planning, next best action, what have you. They’re going to be very specific applications. That innovation has to come from everywhere. And one of the things it needs is clean data. All of these AI applications need clean, concurrent, fast data. So one of the things we did — started about 2 years ago actually is put in a new API on the Vault platform called the Direct Data API, and that was just released this April. 

Veeva’s management has no plans to develop or acquire generative AI solutions currently, but are open to the idea as they observe how the technology evolves; Veeva’s applications do use AI technology, but not specifically generative AI; customers really trust Veeva, so management wants to move carefully when it comes to Veeva developing generative AI applications

We don’t have plans to develop or acquire GenAI solutions today, but that may change in the coming years as we see how GenAI technology evolves, and we determine which use cases can provide consistent value for the industry. In the meantime, we will continue to add advanced automation to our applications. Some, like TMF Bot and RIM Bot, use AI technology, but generally not GenAI…

… We have that trust. We have to continue to earn that trust. So we don’t really get into things that are too speculative. We definitely don’t overpromise. The trust is the most valuable thing we have. So we’ll be really targeted when we get into an AI application if we do. It will be an area where, hey, that’s a use case that we’re pretty sure that can be solved by GenAI, and there’s not a great partner to do it. Okay. Then we might step in because we do have that trusted position.

Veeva’s management lowered the company’s FY2025 revenue guidance slightly (was previously $2.725 billion – $2.74 billion) because of macro challenges and crowding-out from companies wanting to reallocate resources to AI; management is seeing some deferment of spending on core systems because customers are busy investing in AI, but the deferment creates pent-up demand and it’s not spending that has stopped

For fiscal year 2025, we now expect total revenue between $2.700 and $2.710 billion. This is a roughly $30 million reduction compared to our prior guidance, mostly in the services area. As we have said, the macro environment remains challenging as the industry continues to navigate inflation, higher interest rates, global conflicts, political instability, and the Inflation Reduction Act. There is also some disruption in large enterprises as they work through their plans for AI…

…A little more than a year ago, AI really burst upon the scene with GenAI…

…That caused a lot of pressure in our larger enterprises, on the IT department, “Hey, what are we going to do about GenAI? What’s our strategy as a large pharmaceutical company, biotech about AI?” And that we would land in the IT department of these companies. Now for the smaller — our smaller SMB customers, doesn’t land so much. They have other things to think about, other more pertinent, very stressful things. But in the large companies, with tens of thousands of people, they’re looking for these operational efficiencies that they could potentially get through AI and they have a budget to kind of get ahead of that game. So that — by the word disruption, I meant that through a competing priority into our customers, hey, we had some existing plans. Now this AI, we have to plan for what we’re going to do on that. Where are we going to spend on innovation, on experimentation? Who’s going to do that? What budget would we use, that type of thing. So some of that would take an impact onto us, which is core systems. Now those core systems, when we get that type of impact, it will delay a project, but it won’t stop it because these core systems are things you need. You can delay them, but all that does is create somewhat of a pent-up demand.

---

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Adobe, DocuSign, MongoDB, Okta, Salesforce, and Veeva Systems. Holdings are subject to change at any time.

The way I see it, artificial intelligence (or AI), really leapt into the zeitgeist in late-2022 or early-2023 with the public introduction of DALL-E2 and ChatGPT. Both are provided by OpenAI and are software products that use AI to generate art and writing, respectively (and often at astounding quality). Since then, developments in AI have progressed at a breathtaking pace.

With the latest earnings season for the US stock market – for the first quarter of 2024 – coming to its tail-end, I thought it would be useful to collate some of the interesting commentary I’ve come across in earnings conference calls, from the leaders of technology companies that I follow or have a vested interest in, on the topic of AI and how the technology could impact their industry and the business world writ large. This is an ongoing series. For the older commentary:

• 2023 Q1 – here and here
• 2023 Q2 – here and here
• 2023 Q3 – here and here
• 2023 Q4 – here and here

With that, here are the latest commentary, in no particular order:

Airbnb (NASDAQ: ABNB)

Airbnb has been using AI for a long time and has made a lot of progress in the last 12 months, including (1) a computer vision AI model trained with 100 million photos that allows hosts to organise all their photos by room, which leads to higher conversion rates, (2) an AI-powered feature for hosts to reply guests quickly, and (3) a reservation screening technology

We’ve been using AI for a long time. In the last 12 months, we’ve made a lot of progress. I’ll just give you 3 examples of things we’ve done with AI. We made it easier to host. We have a computer vision model that we trained with 100 million photos, and that allows hosts to — like the AI model to organize all their photos by room. Why would you want to do this? Because this increases conversion rate when you do this. We launched last week AI-powered quick replies for hosts. So basically, predicts the right kind of question or answer for a host to pre-generate to provide to guests. And this has been really helpful. And then we’ve made a really big impact on reducing partners in Airbnb with a reservation screening technology.

Airbnb’s management is going much bigger on generative AI; management thinks the biggest near-term impact generative AI can have on Airbnb’s business is in customer service; management thinks that generative AI in the realm of customer service can benefit Airbnb a lot more than hotels and online travel agents (OTAs); AI can solve difficult customer service challenges for Airbnb

So now we’re going much bigger on generative AI. I think I think we’re going to see the biggest impact is going to be on customer service in the near term. I think more than hotels, probably even more than OTA, Airbnb will benefit from generative AI. And the reason why, it’s just a simple structural reason. We have the most like buried inventory. We don’t have any SKUs, and we’re an incredibly global platform. So it’s a very difficult customer service challenge. But imagine an AI agent that can actually like read a corpus of 1,000 pages of policies and be able to help adjudicate and help a customer service agent help a guest from Germany staying with a host in Japan. It’s a very difficult problem that AI can really supplement. 

Airbnb’s management wants to bring AI capabilities from customer service to search and to the broader experience; the end game is to provide an AI-powered concierge

Over time, we’re going to bring the AI capabilities from customer service to search and to the broader experience. And the end game is to provide basically an AI-powered concierge. 

Alphabet (NASDAQ: GOOG)

Alphabet’s management gave a reminder that Alphabet has been an AI-first company since 2016; Alphabet started building TPUs (tensor processing units) in 2016

We’ve been an AI-first company since 2016, pioneering many of the modern breakthroughs that power AI progress for us and for the industry…

… You can imagine we started building TPUs in 2016, so we’ve definitely been gearing up for a long time.

Alphabet’s management rolled out Gemini 1.5 Pro in February, a foundational AI model which has a breakthrough in long context understanding and multimodal capabilities; Gemini 1.5 Pro has been embraced by developers and enterprise customers in a wide range of use cases

In February, we rolled out Gemini 1.5 Pro, which shows dramatic performance enhancements across a number of dimensions. It includes a breakthrough in long context understanding, achieving the longest context window of any large-scale foundation model yet. Combining this with Gemini’s native multimodal understanding across audio, video, text code and more, it’s highly capable. We are already seeing developers and enterprise customers enthusiastically embrace Gemini 1.5 and use it for a wide range of things.

Alphabet’s management thinks that the company has the best infrastructure for AI; Gemini’s training and inference is done with Alphabet’s custom TPU (tensor processing unit) chips; Google Cloud offers the latest generation of Nvidia GPUs (graphics processing units) and Alphabet’s own TPUs

We have the best infrastructure for the AI era… Our data centers are some of the most high-performing, secure, reliable, and efficient in the world. They’ve been purpose-built for training cutting-edge AI models and designed to achieve unprecedented improvements in efficiency. We have developed new AI models and algorithms that are more than 100x more efficient than they were 18 months ago. Our custom TPUs, now in their fifth generation, are powering the next generation of ambitious AI projects. Gemini was trained on and is served using TPUs…

…We offer an industry-leading portfolio of NVIDIA GPUs along with our TPUs. This includes TPU v5p, which is now generally available and NVIDIA’s latest generation of Blackwell GPUs. 

Alphabet’s management is seeing generative AI cause a shift in what people can do with Search, and they think this will lead to a new stage of growth, similar to the outcomes of prior shifts in Search; Alphabet has been experimenting with SGE (Search Generative Experience) for over a year and the company is now bringing AI overseas to main Search results; Alphabet has served billions of queries with its generative AI features; people who use the AI overviews in Google Search increase their search usage and report higher satisfaction with search results; ads that are above or below SGE results were found by users to be helpful; management is confident that SGE with ads will remain relevant; management thinks that the use of generative AI can help Google answer more complex questions and expand the type of queries it can serve

We have been through technology shifts before, to the web, to mobile, and even to voice technology. Each shift expanded what people can do with Search and led to new growth. We are seeing a similar shift happening now with generative AI. For nearly a year, we’ve been experimenting with SGE in search labs across a wide range of queries. And now we are starting to bring AI overviews to the main Search results page. We are being measured in how we do this, focusing on areas where gen AI can improve the search experience while also prioritizing traffic to websites and merchants. We have already served billions of queries with our generative AI features. It’s enabling people to access new information, to ask questions in new ways and to ask more complex questions. Most notably, based on our testing, we are encouraged that we are seeing an increase in search usage among people who use the new AI overviews as well as increased user satisfaction with the results…

…We shared in March how folks are finding ads either above or below the SGE results helpful. We’re excited to have a solid baseline to keep innovating on and confident in the role SGE, including Ads, will play in delighting users and expanding opportunities to meet user needs…

… I think with generative AI in Search, with our AI overviews, I think we will expand the type of queries we can serve our users. We can answer more complex questions as well as, in general, that all seems to carry over across query categories. Obviously, it’s still early, and we are going to be measured and put user experience at the front, but we are positive about what this transition means…

…On SGE in Search, we are seeing early confirmation of our thesis that this will expand the universe of queries where we are able to really provide people with a mix of actual answers linked to sources across the Web and bring a variety of perspectives, all in an innovative way. 

The cost of producing SGE responses has decreased by 80% from when SGE was first introduced a year ago because of work Alphabet has done on its Gemini models and TPUs

A number of technical breakthroughs are enhancing machine speed and efficiency, including the new family of Gemini models and a new generation of TPUs. For example, since introducing SGE about a year ago, machine costs associated with SGE responses have decreased 80% from when first introduced in Labs driven by hardware, engineering, and technical breakthroughs.

Alphabet’s immense reach – 6 products with >2 billion monthly users each, and 15 products with 0.5 billion users – is helpful in distributing AI to users; management has brought AI features to many Alphabet products

We have 6 products with more than 2 billion monthly users, including 3 billion Android devices. 15 products have 0.5 billion users, and we operate across 100-plus countries. This gives us a lot of opportunities to bring helpful gen AI features and multimodal capabilities to people everywhere and improve their experiences. We have brought many new AI features to Pixel, Photos, Chrome, Messages and more. We are also pleased with the progress we are seeing with Gemini and Gemini Advanced through the Gemini app on Android and the Google app on iOS.

Alphabet’s management thinks the company has a clear path to monetisation of AI services through ads, cloud, and subscriptions; Alphabet introduced Gemini Advance, a subscription service to access the most advanced Gemini model, was introduced in 2024 Q1

We have clear paths to AI monetization through Ads and Cloud as well as subscriptions…

… Our Cloud business continues to grow as we bring the best of Google AI to enterprise customers and organizations around the world. And Google One now has crossed 100 million paid subscribers, and in Q1, we introduced a new AI premium plan with Gemini Advanced.

Established enterprises are using Google Cloud for their AI needs (For example: (1) Discover Financial has begun deploying generative AI tools to its 10,000 call center agents, (2) McDonald’s is using gen AI to enhance its customer and employee experiences, and (3) WPP is integrating with Gemini models); more than 60% of funded generative AI (gen AI) start-ups and nearly 90% of gen AI unicorns are also using Google Cloud; more than 1 million developers are now using Alphabet’s generative AI tools; customers can now also ground their generative AI with Google Search and their own data 

At Google Cloud Next, more than 300 customers and partners spoke about their generative AI successes with Google Cloud, including global brands like Bayer, Cintas, Mercedes-Benz, Walmart and many more…

…Today, more than 60% of funded gen AI start-ups and nearly 90% of gen AI unicorns are Google Cloud customers. And customers like PayPal and Kakao Brain are choosing our infrastructure… 

……On top of our infrastructure, we offer more than 130 models, including our own models, open source models and third-party models. We made Gemini 1.5 Pro available to customers as well as Imagine 2.0 at Cloud Next. And we shared that more than 1 million developers are now using our generative AI across tools, including AI Studio and Vertex AI. We spoke about how customers like Bristol-Myers Squibb and Etsy can quickly and easily build agents and connect them to their existing systems. For example, Discover Financial has begun deploying gen AI-driven tools to its nearly 10,000 call center agents to achieve faster resolution times for customers. Customers can also now ground their gen AI with Google Search and their own data from their enterprise databases and applications. In Workspace, we announced that organizations like Uber, Pepperdine University and PennyMac are using Gemini and Google Workspace, our AI-powered agent that’s built right into Gmail, Docs sheets and more…

…To help McDonald’s build the restaurant of the future, we’re deepening our partnership across cloud and ads. Part of this includes them connecting Google Cloud’s latest hardware and data technologies across restaurants globally and starting to apply Gen AI to enhance its customer and employee experiences. Number two, WPP. At Google Cloud Next, we announced a new collaboration that will redefine marketing through the integration of our Gemini models with WPP Open, WPP’s AI-powered marketing operating system, already used by more than 35,000 of its people and adopted by key clients, including The Coca-Cola Company, L’Oreal and Nestle. We’re just getting started here and excited about the innovation this partnership will unlock. 

Alphabet’s management has AI solutions to help advertisers with predicting ad conversions and to match ads with relevant searches; management thinks Alphabet’s ability to help advertisers find customers and grow their advertising ROI (return on investment) is getting better as the company’s AI models improve

We’ve talked about whole solutions like Smart Bidding use AI to predict future ad conversions and their value in helping businesses stay agile and responsive to rapid shifts in demand and how products like broad match leverage LLMs to match ads to relevant searches and help advertisers respond to what millions of people are searching for…

…As advances accelerate in our underlying AI models, our ability to help businesses find users at speed and scale and drive ROI just keeps getting better.

Alphabet’s management introduced Gemini into Performance Max (PMax) in February and early results show PMax users are 63% more likely to publish a campaign with good or excellent ad strength and those who improve their ad strength on PMax to excellent see a 6% increase in conversions; PMax is available to all US advertisers and is starting to be rolled out internationally

In February, we rolled Gemini into PMax. It’s helping curate and generate text and image assets so businesses can meet PMax asset requirements instantly. This is available to all U.S. advertisers and starting to roll out internationally in English, and early results are encouraging. Advertisers using PMax asset generation are 63% more likely to publish a campaign with good or excellent ad strength. And those who improve their PMX ad strength to excellent see 6% more conversions on average.

Advertisers who use Alphabet’s ACA (automatically created assets) feature that is powered by generative AI see conversions increase by 5%

We’re also driving improved results for businesses opting into automatically created assets, which are supercharged with gen AI. Those adopting ACA see, on average, 5% more conversions at a similar cost per conversion in Search and Performance Max campaigns.

Alphabet’s Demand Gen AI-powered service helps advertisers engage with new and existing customers across Youtube, Shorts, Gmail, and Discover; movie studio Lionsgate tested Demand Gen for a movie’s promotion and saw that it provided an 85% more efficient CPC (cost per click) and 96% more efficient cost per page view compared to social benchmarks; Lionsgate has used Demand Gen for two more films; Alphabet recently introduced new tools in Demand Gen

And then there’s Demand Gen. Advertisers are loving its ability to engage new and existing customers and drive purchase consideration across our most immersive and visual touch points like YouTube, Shorts, Gmail and Discover. Hollywood film and TV studio, Lionsgate, partnered with Horizon Media to test what campaign type will deliver the most ticketing page views for its The Hunger Games: Ballad of Songbirds and Snakes film. Over a 3-week test, demand gen was significantly more efficient versus social benchmarks with an 85% more efficient CPC and 96% more efficient cost per page view. Lionsgate has since rolled out Demand Gen for 2 new titles. We’re also bringing new creative features to demand gen. Earlier this month, we announced new generative image tools to help advertisers create high-quality assets in a few steps with a few simple prompts. This will be a win for up-leveling visual storytelling and testing creative concepts more efficiently.

Google Cloud had 28% revenue growth in 2024 Q1 (was 26% in 2023 Q4), driven by an increasing contribution from AI; management sees the growth of Google Cloud being underpinned by the benefits AI provides for customers, and management wants to invest aggressively in cloud while remaining focused on profitable growth; Alphabet’s big jump capex in 2024 Q1 (was $6.3 billion in 2023 Q1) was mostly for technical infrastructure and reflects management’s confidence in the opportunities offered by AI; management expects Alphabet’s quarterly capex for the rest of 2024 to be similar to what was seen in 2024 Q1; management has no view on 2025 capex at the moment; management sees Google Cloud hitting an inflection point because of AI

Turning to the Google Cloud segment. Revenues were $9.6 billion for the quarter, up 28%, reflecting significant growth in GCP with an increasing contribution from AI and strong Google Workspace growth, primarily driven by increases in average revenue per seat. Google Cloud delivered operating income of $900 million and an operating margin of 9%…

…With respect to Google Cloud, performance in Q1 reflects strong demand for our GCP infrastructure and solutions as well as the contribution from our Workspace productivity tools. The growth we are seeing across Cloud is underpinned by the benefit AI provides for our customers. We continue to invest aggressively while remaining focused on profitable growth…

…With respect to CapEx, our reported CapEx in the first quarter was $12 billion, once again driven overwhelmingly by investment in our technical infrastructure, with the largest component for servers followed by data centers. The significant year-on-year growth in CapEx in recent quarters reflects our confidence in the opportunities offered by AI across our business. Looking ahead, we expect quarterly CapEx throughout the year to be roughly at or above the Q1 level, keeping in mind that the timing of cash payments can cause variability in quarterly reported CapEx…

…And then with respect to 2025, as you said, it’s premature to comment so nothing to add on that…

…On the Cloud side, obviously, it’s definitely a point of inflection overall. I think the AI transformation is making everyone think about their whole stack, and we are engaged in a number of conversations. I think paid AI infrastructure, people are really looking to Vertex AI, given our depth and breadth of model choice, or using Workspace to transform productivity in your workplace, et cetera. So I think the opportunities there are all related to that, both all the work we’ve built up and AI being a point of inflection in terms of driving conversations. I think you’ll see us do it both organically and with a strong partner program as well. So we’ll do it with a combination.

Alphabet’s management thinks the AI transition is a once-in-a-generation opportunity; it’s the first time they think Alphabet can work on AI in a horizontal way

I think the AI transition, I think it’s a once-in-a-generation kind of an opportunity. We’ve definitely been gearing up for this for a long time. You can imagine we started building TPUs in 2016, so we’ve definitely been gearing up for a long time…

… The real opportunities we see is the scale of research and innovation, which we have built up and are going to continue to deliver. I think for the first time, we can work on AI in a horizontal way and it impacts the entire breadth of the company, be it Search, be it YouTube, be it Cloud, be it Waymo and so on. And we see a rapid pace of innovation in that underlying.

Alphabet’s management thinks that, with regards to monetising the opportunity of smartphone-based AI searches, there will be search use-cases that can be fulfilled on-device, but there will be many, many search use-cases that will require the internet

[Question] As users start searching on smartphones and those searches are basically rendered on the model, on the phone, without accessing the web, how do you guys anticipate monetizing some of these smartphone-based behaviors that are kind of run on the edge?

[Answer] If you look at what users are looking for, people are looking for information and an ability to connect with things outside. So I think there will be a set of use cases which you will be able to do on device. But for a lot of what people are looking to do, I think you will need the richness of the cloud, the Web and you have to deliver it to users. So again, to my earlier comments, I think through all these moments, you saw what we have done with Samsung with Circle to Search. I think it gives a new way for people to access Search conveniently wherever they are. And so we view this as a positive way to bring our services to users in a more seamless manner. So I think it’s positive from that perspective. In terms of on-device versus cloud, there will be needs which can be done on-device and we should to help it from a privacy standpoint. But there are many, many things for which people will need to reach out to the cloud. And so I don’t see that as being a big driver in the on-cloud versus off-cloud in any way.

Amazon (NASDAQ: AMZN)

Amazon’s management recently launched a new generative AI tool for third-party sellers to quickly create product detail pages on Amazon using just the sellers’ URL to their websites; more than 100,000 third-party sellers on Amazon are already using at least one of Amazon’s generative AI tools

We’ve recently launched a new generative AI tool that enables sellers to simply provide a URL to their own website, and we automatically create high-quality product detail pages on Amazon. Already, over 100,000 of our selling partners have used one or more of our gen AI tools. 

Amazon’s management is seeing AWS customers being excited about leveraging generative AI to change their customer experiences and businesses; AWS’s AI business is already at a multibillion-dollar revenue rate; AWS AI’s business is driven by a few things, including the fact that many companies are still building their models; management expects more models to be built on AWS over time because of the depth of AI offerings AWS has

Our AWS customers are also quite excited about leveraging gen AI to change the customer experiences and businesses. We see considerable momentum on the AI front where we’ve accumulated a multibillion-dollar revenue run rate already…

… I mentioned we have a multibillion-dollar revenue run rate that we see in AI already, and it’s still relatively early days. And I think that there’s — at a high level, there’s a few things that we’re seeing that’s driving that growth. I think first of all, there are so many companies that are still building their models. And these range from the largest foundational model builders like Anthropic, you mentioned, to every 12 to 18 months or building new models. And those models consume an incredible amount of data with a lot of tokens, and they’re significant to actually go train. And a lot of those are being built on top of AWS, and I expect an increasing amount of those to be built on AWS over time because our operational performance and security and as well as our chips, both what we offer from NVIDIA. But if you take Anthropic, as an example, they’re training their future models on our custom silicon on Trainium. And so I think we’ll have a real opportunity for a lot of those models to run on top of AWS.

Amazon’s management’s framework for thinking about generative AI consists of 3 layers –  the first is the compute layer, the second is LLMs as a service, the third is the applications that run on top of LLMs – and Amazon continues to add capabilities in all 3

You heard me talk about our approach before, and we continue to add capabilities at all 3 layers of the gen AI stack. At the bottom layer, which is for developers and companies building models themselves, we see excitement about our offerings…

…The middle layer of the stack is for developers and companies who prefer not to build models from scratch but rather seek to leverage an existing large language model, or LLM, customize it with their own data and have the easiest and best features available to deploy secure high-quality, low-latency, cost-effective production gen AI apps…

…The top of the stack are the gen AI applications being built. 

Amazon’s management thinks AWS has the broadest selection of Nvidia compute instances but also sees high demand for Amazon’s custom silicon, Trainium and Inferentia, as they provide favourable price performance benefits; larger quantities of Amazon’s latest Trainium chip, Trainium 2, will arrive in 2024 H2 and early 2025; Anthropic’s future models will be trained on Tranium

We have the broadest selection of NVIDIA compute instances around, but demand for our custom silicon, Trainium and Inferentia, is quite high given its favorable price performance benefits relative to available alternatives. Larger quantities of our latest generation Trainium2 is coming in the second half of 2024 and early 2025…

…But if you take Anthropic, as an example, they’re training their future models on our custom silicon on Trainium. 

SageMaker, AWS’s fully-managed machine learning service, has helped (1) Perplexity AI train models 40% faster, (2) Workday reduce inference latency by 80%, and (3) NatWest reduce time to value for AI from 12-18 months to less than 7 months; management is seeing an increasing number of AI model builders standardising on SageMaker

Companies are also starting to talk about the eye-opening results they’re getting using SageMaker. Our managed end-to-end service has been a game changer for developers in preparing their data for AI, managing experiments, training models faster, lowering inference latency, and improving developer productivity. Perplexity.ai trains models 40% faster than SageMaker. Workday reduces inference latency by 80% with SageMaker, and NatWest reduces its time to value for AI from 12 to 18 months to under 7 months using SageMaker. This change is how challenging it is to build your own models, and we see an increasing number of model builders standardizing on SageMaker.

Amazon’s management thinks Amazon Bedrock, a LLM-as-a-service offering, has the broadest selection of LLMs (large language models) for customers in addition to retrieval augmented generation (RAG) and other features; Bedrock offers high-profile LLMs – such as Anthropic’s Claude 3 and Meta’s Llama 3 – in addition to Amazon’s own Titan models; Custom Model Import is a new feature from Bedrock that satisfies a customer request (the ability to import models from SageMaker or elsewhere into Bedrock in a simple manner) that nobody has yet met; management is seeing customers being excited about Custom Model Import; Bedrock has tens of thousands of customers

 This is why we built Amazon Bedrock, which not only has the broadest selection of LLMs available to customers but also unusually compelling model evaluation, retrieval augmented generation, or RAG, to expand model’s knowledge base, guardrails to safeguard what questions applications will answer, agents to complete multistep tasks, and fine-tuning to keep teaching and refining models. Bedrock already has tens of thousands of customers, including adidas, New York Stock Exchange, Pfizer, Ryanair and Toyota. In the last few months, Bedrock’s added Anthropic’s Claude 3 models, the best-performing models in the planet right now; Meta’s Llama 3 models; Mistral’s various models, Cohere’s new models and new first-party Amazon Titan models.

A week ago, Bedrock launched a series of other features, but perhaps most importantly, Custom Model Import. Custom Model Import is a sneaky big launch as it satisfies a customer request we’ve heard frequently and that nobody has yet met. As increasingly more customers are using SageMaker to build their models, they’re wanting to take advantage of all the Bedrock features I mentioned earlier that make it so much easier to build high-quality production-grade gen AI apps. Bedrock Custom Model Import makes it simple to import models from SageMaker or elsewhere into Bedrock before deploying their applications. Customers are excited about this, and as more companies find they’re employing a mix of custom-built models along with leveraging existing LLMs, the prospect of these 2 linchpin services in SageMaker and Bedrock working well together is quite appealing…

…And the primary example we see there is how many companies, tens of thousands of companies, already are building on top of Amazon Bedrock.

Amazon’s management has announced the general availability of Amazon Q, a highly-capable generative AI-powered assistant; Amazon Q helps developers generate code, test code, debug code, and can save developers months of work when moving from older versions of Java to newer ones; Amazon Q has an Agents capability which can autonomously perform a range of tasks, including (1) implementing application features, and (2) parsing a company’s entire data stock to create summaries and surface insights; Amazon Q also has Q Apps, which lets employees describe in natural language what app they want to build on top of internal data; management believes that Q is the most functionally-capable AI-powered assistant for software development and data, as Q outperforms competitors; many companies are already using Amazon Q

And today, we announced the general availability of Amazon Q, the most capable generative AI-powered assistant for software development and leveraging company’s internal data.

On the software development side, Q doesn’t just generate code. It also tests code, debugs coding conflicts, and transforms code from one form to another. Today, developers can save months using Q to move from older versions of Java to newer, more secure and capable ones. In the near future, Q will help developers transform their .NET code as well, helping them move from Windows to Linux.

Q also has a unique capability called Agents, which can autonomously perform a range of tasks, everything from implementing features, documenting, and refactoring code to performing software upgrades. Developers can simply ask Amazon Q to implement an application feature such as asking it to create an add to favorites feature in a social sharing app, and the agent will analyze their existing application code and generate a step-by-step implementation plan, including code changes across multiple files and suggested new functions. Developers can collaborate with the agent to review and iterate on the plan, and then the agent implements it, connecting multiple steps together and applying updates across multiple files, code blocks and test suites. It’s quite handy. On the internal data side, most companies have large troves of internally relevant data that resides in wikis, Internet pages, Salesforce, storage repositories like Amazon S3 and a bevy of other data stores and SaaS apps that are hard to access. It makes answering straightforward questions about company policies, products, business results, code, people, and many other topics hard and frustrating. Q makes this much simpler. You can point Q at all of your enterprise data repositories and it will search all this data, summarize logically, analyze trends, engage in dialogue with customers about this data.

We also introduced today a powerful new capability called Q Apps, which lets employees describe a natural language what apps they want to build on top of this internal data and Q Apps will quickly generate that app. This is going to make it so much easier for internal teams to build useful apps from their own data.

Q is not only the most functionally capable AI-powered assistant for software development and data but also setting the standard for performance. Q has the highest-known score and acceptance rate for code suggestions, outperforms all other publicly benchmarkable competitors and catching security vulnerabilities, and leads all software development assistants on connecting multiple steps together and applying automatic actions. Customers are gravitating to Q, and we already see companies like Brightcove, British Telecom, Datadog, GitLab, GoDaddy, National Australia Bank, NCS, Netsmart, Slam, Smartsheet, Sun Life, Tata Consultancy Services, Toyota, and Wiz using Q, and we’ve only been in beta until today.

Amazon’s management believes that AWS has a meaningful edge in security elements when it comes to generative AI, and this has led to companies moving their AI focus to AWS

I’d also caution folks not to overlook the security and operational performance elements of these gen AI services. It’s less sexy but critically important. Most companies care deeply about the privacy of the data in their AI applications and the reliability of their training and production apps. If you’ve been paying attention to what’s been happening in the last year or so, you can see there are big differences between providers on these dimensions. AWS has a meaningful edge, which is adding to the number of companies moving their AI focus to AWS.

Amazon’s management sees Amazon’s capex increasing meaningfully in 2024 compared to 2023 ($48.4 billion in 2023) because of AWS’s accelerating growth and high demand for generative AI; the capex in 2024 will go mostly towards technology infrastructure; the capex of $14 billion in 2024 Q1 will be the low quarter for the year;

We expect the combination of AWS’ reaccelerating growth and high demand for gen AI to meaningfully increase year-over-year capital expenditures in 2024, which given the way the AWS business model works is a positive sign of the future growth…

…As a reminder, we define these as the combination of CapEx plus equipment finance leases. In 2023, overall capital investments were $48.4 billion…

…We do see, though, on the CapEx side that we will be meaningfully stepping up our CapEx and the majority of that will be in our — to support AWS infrastructure and specifically generative AI efforts…

…We’re talking about CapEx. Right now, in Q1, we had $14 billion of CapEx. We expect that to be the low quarter for the year.

Amazon’s management is very bullish on AWS, as 85% or more of global IT spend remains on-premise, even though AWS is already at at $100 billion-plus revenue run rate; in addition, there’s demand for generative AI, most of which will be created in the next few decades from scratch and on the cloud

We remain very bullish on AWS. We’re at $100 billion-plus annualized revenue run rate, yet 85% or more of the global IT spend remains on-premises. And this is before you even calculate gen AI, most of which will be created over the next 10 to 20 years from scratch and on the cloud. There is a very large opportunity in front of us. 

Amazon’s management thinks the generative AI opportunity is something they have not seen since the cloud or internet

We have a lot of growth in front of us, and that’s before the generative AI opportunity, which I don’t know if any of us have seen a possibility like this in technology in a really long time, for sure, since the cloud, perhaps since the Internet. 

Amazon’s management thinks much more money will be spent on AI inference than on model training; management sees quite a few companies that are building their generative AI applications to do inference on AWS

I think the thing that people sometimes don’t realize is that while we’re in the stage that so many companies are spending money training models, once you get those models into production, which not that many companies have, but when you think about how many generative AI applications will be out there over time, most will end up being in production when you see the significant run rates. You spend much more in inference than you do in training because you train only periodically, but you’re spinning out predictions and inferences all the time. And so we also see quite a few companies that are building their generative AI applications to do inference on top of AWS.

Amazon’s management sees both training and inference being really big drivers for AWS; this is helped by the fact that these AI models will work with companies’ data and the security surrounding the data is important for companies, and AWS has a meaningful edge in security

We see both training and inference being really big drivers on top of AWS. And then you layer on top of that the fact that so many companies, their models and these generative AI applications are going to have their most sensitive assets and data. And it’s going to matter a lot to them what kind of security they get around those applications. And yes, if you just pay attention to what’s been happening over the last year or 2, not all the providers have the same track record. And we have a meaningful edge on the AWS side so that as companies are now getting into the phase of seriously experimenting and then actually deploying these applications to production, people want to run their generative AI on top of AWS.

Apple (NASDAQ: AAPL)

Apple’s management continues to feel bullish about Apple’s opportunity in generative AI; Apple is making significant investments in the area and will be sharing details soon; management thinks Apple has advantages with AI given its unique combination of hardware, software, services, custom silicon (with industry-leading neural engines), and privacy

We continue to feel very bullish about our opportunity in generative AI. We are making significant investments, and we’re looking forward to sharing some very exciting things with our customers soon. We believe in the transformative power and promise of AI, and we believe we have advantages that will differentiate us in this new era, including Apple’s unique combination of seamless hardware, software, and services integration, groundbreaking Apple silicon with our industry-leading neural engines, and our unwavering focus on privacy, which underpins everything we create. 

Apple’s management does not expect Apple’s capex to inflect higher, nor the composition of the capex to change much, even as the company leans into AI

[Question] As Apple leans more into AI and generative AI, should we expect any changes to the historical CapEx cadence that we’ve seen in the last few years of about $10 billion to $11 billion per year? Or any changes to how we may have historically thought about the split between tooling, data center, and facilities?

[Answer]  We are obviously very excited about the opportunity with GenAI. We obviously are pushing very hard on innovation on every front, and we’ve been doing that for many, many years. Just during the last 5 years, we spent more than $100 billion in research and development. As you know, on the CapEx front, we have a bit of a hybrid model where we make some of the investments ourselves. In other cases, we share them with our suppliers and partners. On the manufacturing side, we purchase some of the tools and manufacturing equipment. In some of the cases, our suppliers make the investment. And we do something similar on the data center side. We have our own data center capacity and then we use capacity from third parties. It’s a model that has worked well for us historically, and we plan to continue along the same lines going forward.

Apple’s management will soon share their thoughts on how Apple intends to monetise AI on its devices – but not today

[Question] You’ve obviously mentioned your excitement around generative AI multiple times. I’m just curious how Apple is thinking about the different ways in which you can monetize this technology because, historically, software upgrades haven’t been a big factor in driving product cycles. And so could AI be potentially different?

[Answer] I don’t want to get in front of our announcements obviously. I would just say that we see generative AI as a very key opportunity across our products, and we believe that we have advantages that set us apart there. And we’ll be talking more about it in — as we go through the weeks ahead.

Arista Networks (NYSE: ANET)

Arista Networks’ management sees an addressable market of US$60 billion in client-to-cloud AI networking

Amidst all the network consolidation, Arista is looking to establish ourselves as the pure-play networking innovator, for the next era, addressing at least a $60 billion TAM in data-driven client-to-cloud AI networking.

Arista Networks’ management is pleased with the momentum they are seeing in the company’s customer segments, including the Cloud and AI Titans segment; management is becoming increasingly constructive about hitting their 2025 target of US$750 million in AI revenue; the 2025 target of US$750 million is not a hockey-stick target, but a glide path

We are quite pleased with the momentum across all our 3 sectors: Cloud and AI Titans, Enterprise and Providers. Customer activity is high as Arista continues to impress our customers and prospects with our undeniable focus on quality and innovation…

… A good AI network needs a good data strategy, delivered by our highly differentiated EOS and network data lake architecture. We are, therefore, becoming increasingly constructive about achieving our AI target of $750 million in 2025…

…When you think about the $750 million target that has become more constructive to Jayshree’s prepared remarks, that’s a glide path. So it’s not 0 in ’24, It’s a glide path to ’25. 

Traditional networking discards data as the network changes state, but recent developments in AI show how important it is to gather and store large data sets – this is a problem Arista Networks’ management is solving through the company’s NetDL (Network Data Lake) platform, which streams every piece of network data in real time and archives the full data history

From the inception of networking decades ago, networking has involved rapidly changing data. Data about how the network is operating, which paths through the network our best and how the network is being used. But historically, most of this data was to simply discarded as the network changes state and that which was collected can be difficult to interpret because it lacks context. Network addresses and port numbers by themselves, provide a little insight into what users are doing or experiencing.

Recent developments in AI have proved the value of data. But to take advantage of these breakthroughs, you need to gather and store large data sets, labeled suitably for machine learning. Arista is solving this problem with NetDL, we continually monitor every device, not simply taking snapshots, but rather streaming every network event, every counter, every piece of data in real time, archiving a full history in NetDL. Alongside this device data, we also collect flow data and inbound network telemetry data gathered by our switches. Then we enrich this performance data further with user, service and application layer data from external sources outside the network, enabling us to understand not just how each part of the network is performing, but also which users are using the network for what purposes. And how the network behavior is influencing their experience. NetDL is a foundational part of the EOS stack, enabling advanced functionality across all of our use cases. For example, in AI fabrics, NetDL enables fabric-wide visibility, integrating network data and NIC data to enable operators to identify misconfigurations or misbehaving hosts and pinpoint performance bottlenecks.

Any slowdown in the network when running generative AI training tasks can reduce processor performance by 30% or more

As generative AI training tasks evolve, they are made up of many thousands of individual iterations. Any slowdown due to network and critically impact the application performance, creating inefficient wait stage and idling away processor performance by 30% or more. The time taken to reach coherence known, as job completion time is an important benchmark achieved by building proper scale-out AI networking to improve the utilization of these precious and expensive GPUs. 

A Cloud and AI Titan customer of Arista Networks used the company’s product to build a 24,000 node GPU cluster for complex AI training tasks; Arista Networks’ product offered an improvement of at least 10% on job completion performance across all packet sizes versus InfiniBand; in Arista Networks’ four recent AI Ethernet clusters that was won versus InfiniBand, management is seeing all four projects migrate from trials to pilots; Arista Networks will be connecting thousands of GPUs in the four projects this year and management expects to connect 10,000 to 100,000 GPUs in 2025; ethernet was traditionally considered to have loss properties while InfiniBand was traditionally considered to be lossless, but when ethernet is used in actual GPU clusters, ethernet is 10% faster than Infiniband; management expects improvement in ethernet’s performance relative to Infiniband in the future, driven partly by the Ultra Ethernet Consortium 

In a recent blog from one of our large Cloud and AI Titan customers, Arista was highlighted for building a 24,000 node GPU cluster based on our flagship 7800 AI Spine. This cluster tackles complex AI training tasks that involve a mix of model and data parallelization across thousands of processors and ethernet is proving to offer at least 10% improvement of job completion performance across all packet sizes versus InfiniBand…

…If you recall, in February, I shared with you that we are progressing well in 4 major AI Ethernet clusters, that we won versus InfiniBand recently. In all 4 cases, we are now migrating from trials to pilots, connecting thousands of GPUs this year, and we expect production in the range of 10,000 to 100,000 GPUs in 2025…

…Historically, as you know, when you look at InfiniBand and Ethernet in isolation, there are a lot of advantages of each technology. Traditionally, InfiniBand has been considered lossless and Ethernet is considered to have some loss properties. However, when you actually put a full GPU cluster together along with the optics and everything, and you look at the coherents of the job completion time across all packet sizes, data has shown that and this is data that we have gotten from third parties, including Broadcom, that just about in every packet size in a real-world environment independent of the comparing those technologies, the job completion time of Ethernet was approximately 10% faster. So you can look at these things in silos. You can look at it in a practical cluster and in a practical cluster we are already seeing improvements on Ethernet. Now don’t forget, this is just Ethernet as we know it today. Once we have the Ultra Ethernet Consortium and some of the improvements you’re going to see on packet spring and dynamic load balancing and congestion control, I believe those numbers will get even better. 

Arista Networks’ management is witnessing an inflection of AI networking and expects the trend to continue both in the short and long run; management is seeing ethernet emerging as critical infrastructure for both front-end and back-end AI data centers; AI applications require seamless communication between the front-end (includes CPUs, or central processing units) and back-end (includes GPUs and AI accelerators); management is seeing ethernet at scale becoming the de facto network and premium choice for scaled-out AI training workloads

We are witnessing an inflection of AI networking and expect this to continue throughout the year and decade. Ethernet is emerging as a critical infrastructure across both front-end and back-end AI data centers. AI applications simply cannot work in isolation and demand seamless communication among the compute nodes, consisting of back-end GPUs and AI accelerators and as well as the front end nodes like the CPUs, alongside storage and IP/WAN systems as well…

…Ethernet at scale is becoming the de facto network at premier choice for scale-out AI training workloads.

Arista Networks’ management thinks that visibility on new AI and cloud projects is getting better and has now improved to at least 6 months

In summary, as we continue to set the direction of Arista 2.0 networking, our visibility to new AI and cloud projects is improving and our enterprise and provider activity continues to progress well…

…In the Cloud and AI Titans in November, we were really searching for even 3 months visibility, 6 would have been amazing. Today, I think after a year of tough situations for us where the Cloud Titans were pivoting rather rapidly to AI and not thinking about the Cloud as much. We’re now seeing a more balanced approach where they’re still doing AI, which is exciting, but they’re also expanding their regions on the Cloud. So I would say our visibility has now improved to at least 6 months and maybe it gets longer as time goes by.

Arista Networks’ management still sees Infiniband as the de facto network of choice for AI workloads, but ethernet is gaining ground; management sees ethernet as being the eventual winner against InfiniBand because ethernet has a long history of 50 years that gives it an advantage (Metcalfe’s law) 

And then sometimes we see them, obviously, when they’re pushing InfiniBand, which has been, for most part, the de facto network of choice. You might have heard me say, last year or the year before, I was outside looking into this AI networking. But today, we feel very pleased that we are able to be the scale-out network for NVIDIA’s, GPUs and NICs based on Ethernet.,,

…This InfiniBand topic keeps coming up. And I’d just like to point out that Ethernet is about 50 years old. And over those 50 years, Ethernet has come head-to-head with a bunch of technologies like Token ring, SONET, ATM, FDDI, HIPPI, Scalable Coherent Interconnect, [ Mirrornet ]. And all of these battles have one thing in common. Ethernet won. And the reason why is because of Metcalfe’s law, the value of a network is quadratic in the number of nodes of the interconnect. And so anybody who tries to build something which is not Ethernet, is starting off with a very large quadratic disadvantage. And any temporary advantage they have because of the — some detail of the tech cycle is going to be quickly overwhelmed by the connectivity advantage you have with Ethernet.

Arista Networks’ management does not see Nvidia as a direct competitor for ethernet; management also believes that Arista Networks’ focus and experience are advantages

We don’t see NVIDIA as a direct competitor yet on the Ethernet side. I think it’s 1% of their business. It’s 100% of our business. So we don’t worry about that overlap at all. And we think we’ve got 20 years of founding to now experience to make our Ethernet switching better and better at both on the front end and back end. So we’re very confident that Arista can build the scale up network and work with NVIDIA scale-up GPUs.

Within AI networking, Arista Networks’ management is seeing the first use case emerging to be the build-out of the fastest training workloads and clusters

The first use case that’s emerging for AI networking is, let’s just build the fastest training workloads and clusters. And they’re looking at performance. Power is a huge consideration, the cooling of the GPUs is a huge part of it. You would be surprised to hear a lot of times, it’s just waiting on the facilities and waiting for the infrastructure to be set up, right?

Arista Networks’ management is seeing Tier 2 cloud providers starting to pick up AI initiatives, although the Tier 2 providers are not close to the level the activity as the Cloud Titans

The Tier 2 cloud providers, I want to speak to them for a moment because not only are they strong for us right now, but they are starting to pick up some AI initiatives as well. So they’re not as large as close as the Cloud Titans, but the combination of the Service Providers and the Tier 2 Specialty Providers is also seeing some momentum.

Arista Networks is seeing GPU lead times improve significantly 

The GPU, the number of GPUs, the location of the GPUs, the scale of the GPUs, the locality of these GPUs, should they go with Blackwell should they build with a scale up inside the server or scale out to the network. So the whole center of gravity, what’s nice to watch which is why we’re more constructive on the 2025 numbers is that the GPU lead times have significantly improved, which means more and more of our customers will get more GPUs, which in turn means they can build out to scale our network.

Arista Networks’ management is not seeing any pause in their customers’ investments in GPU clusters and networking just to wait for the delivery of Nvidia’s latest Blackwell AI chips; Arista Networks’ networking products can perform the required networking tasks well regardless of what GPU is used

[Question] I want to go back to AI, the road map and the deployment schedule for Blackwell. So it sounds like it’s a bit slower than maybe initially expected with initial customer delivery late this year. How are you thinking about that in terms of your road map specifically and how that plays into what you’re thinking about ’25 in a little bit more detail. And does that late delivery maybe put a little bit of a pause on maybe some of the cloud spend in the fall of this year as there seems to be somewhat of a technology transition going on towards Blackwell away from the Legacy product?

[Answer] We’re not seeing a pause yet. I don’t think anybody is going to wait for Blackwell necessarily in 2024 because they’re still bringing up their GPU clusters. And how a cluster is divided across multiple tenants, the choice of host, memory, storage architectures, optimizations on the GPU for collective communication, libraries, specific workloads, resilience, visibility, all of that has to be taken into consideration. All this to say, a good scale-out network has to be built, no matter whether you’re connecting to today’s GPUs or future Balckwells. And so they’re not going to pause the network because they’re waiting for Blackwell. they’re going to get ready for the network, whether it connects to a Blackwell or a current H100. So as we see it, the training workloads and the urgency of getting the best job completion time is so important that they’re not going to spare any investments on the network side and the network side can be ready no matter what the GPU is.

ASML (NASDAQ: ASML)

ASML’s management sees no change to the company’s outlook for 2024 from what was mentioned in the 2023 Q4 earnings call, with AI-related applications still driving demand, Memory demand being driven by DRAM technology node transitions to support DDR5 and HBM, and Logic customers digesting capacity additions made in 2023

Looking at the market segments, we see a similar environment as communicated last quarter with demand momentum from AI-related applications. Memory demand is primarily driven by DRAM technology node transitions in support of advanced memories such as DDR5 and HBM. Logic customers continue to digest the significant capacity additions made over the last year — over the past year

ASML’s management sees some macro uncertainties as still being present, but the long-term trends in the company’s business (AI, electrification, energy transition) are intact

There are still some uncertainties. I would say primarily macro uncertainties. That’s still clearly there…

…If you look at the trends in the industry, if you look at, and I’m talking about the cyclicality trends in the industry, so like the utilization going up, inventory downstream being managed to more normal levels. I think it’s pretty clear that the industry is in its upturn and therefore we do believe that by 2024 we’re going to see a recovery. Clearly a recovery of the industry. So then fast forward to 2025. Then what do we find ourselves in? First off, I think we will find ourselves in 2025 in the midst of the upturn. So that’s a positive. Second – and we’ve talked about that many times – the secular trends are really strong. If you look at AI, if you look at electrification, if you look at the energy transition. It’s all very strong, very positive momentum behind it. So the secular trends are very, very strong. That is also something that I think will yield in 2025. Finally, if you just look at all the fab openings that have been indicated by our customers. The recent news on positive outcomes of CHIPS Act money allocation. All of that is very strong, very supportive for new fab openings across the globe. I think by 2025 you will see all three of those coming together. New fab openings, strong secular trends and the industry in the midst of its upturn. So that’s why we’re doing what we’re doing. Which is really preparing for that ramp, for that momentum that we see being built up.

ASML’s management thinks that AI will be driving demand for leading-edge and high-performance compute; AI is itself driven by massive amounts of data and the overlay of smart software over the data; management also thinks that IoT (Internet of Things) will be an area with plenty of AI applications

You’re basically saying what will drive leading-edge, high-performance compute. But you’re absolutely right. I mean, when you think about high-performance compute, and especially in the context of AI, and I’ve said this many, many times before, AI is driven by massive amounts of data and about also understanding the correlation between those data elements and then overlaying that with smart software. But — and I also believe, it’s actually what I’m seeing and what I’m hearing is that IoT in the industrial space will actually be in — will be an area where we will see a lot of AI applications. Well, in order to collect all that data, you need sensors because you’ve got all kinds of examples, whether it’s the car or whether it’s life science, medical equipment, it’s about sensing, and that is really the domain of mainstream semiconductors.

ASML’s management is seeing the software world enjoying 30% to 50% increases in productivity because of the use of AI

And when you think about AI, I mean, some of these examples, and especially in the software space where you see productivity, just the calculated productivity advantages of 30% to 50%, then the value of the next-generation transistor will be huge.

Coupang (NYSE: CPNG)

Coupang’s management is exploring both the company’s own foundational AI models as well as those from third-parties; AI has been a core part of Coupang’s strategy and management has been deploying the technology in many areas of the company’s business; management is excited about the potential of AI, but will be testing its ability to generate returns for the business

On AI, we are exploring, both for us, as you mentioned, foundational models as well as our own. Machine learning and AI continues to be — have been a core part of our strategy. We’ve deployed them in many facets of our business, from supply chain management to same-day logistics. We’re also seeing tremendous potential with large language models in a number of areas from search and ads, to catalog and operations, among others. There’s exciting potential for AI that we see and we see opportunities for it to contribute even more significantly to our business. But like any investment we make, we’ll test and iterate and then invest further only in the cases where we see the greatest potential for return.

Datadog (NASDAQ: DDOG)

Datadog’s management has announced general availability of Bits AI for incident management, where Bits AI can produce auto-generated incident summaries for incident responders

In the MegaGenAI space, we announced general availability of Bits AI for incident management. By using Bits AI for incident management, incident responders get auto-generated incident summaries to quickly understand the context and scope of a complex incident. And users can also enqure Bits AI to ask about related incidents and to form tasks on the fly from incident creation to resolution.

There’s growing interest in AI from Datadog’s customers, and the company’s next-gen AI customers accounted for 3.5% of ARR (was 3% in 2023 Q4); the percentage of ARR from next-gen AI customers is a metric that management thinks will become less relevant over time as AI adoption broadens

We’re also continuing to see more interest in AI from our customers. As a data point, ARR for next GenAI customers was about 3.5% of our total, a strong sign of the growing ecosystem of companies in this area…

…I’m not sure this is a metric we’ll keep bringing up. It was interesting for us to look at this small group of early AI-native companies to get a sense of what might come next in the world of AI. But I think as we — as time goes by and as AI adoption broadens, I think it becomes less and less relevant. 

Datadog has AI integrations that allow customers to pull their AI data into the Datadog platform; around 2,000 customers are already using 1 or more of the AI integrations

To help customers understand AI technologies and bring them into production applications, our AI integrations allow customers to pull their AI data into the Datadog platform. And today, about 2,000 of our customers are using 1 or more of these AI integrations. And we’ve continued to keep up with the rapid innovation in this space. For example, adding a new integration in Q1 with the NVIDIA Triton [indiscernible] server. 

Datadog’s management has announced general availability for Event Management in the cloud service management area; Event Management reduces the volume of alerts and events Datadog’s customers have to deal with; with Event Management, Datadog now has a full AI solution that helps teams automate remediation, proactively prevent outages and reduce the impact of incidents.

In the cloud service management area, we released event management in the general availability. Our customers face increasing complexity at scale, causing the volume of alerts and events to explode, which makes it difficult for teams to identify, prioritize, summarize and route issue to the right responders. Event management addresses this challenge by automatically reducing a massive volume of events and alerts into actionable insights. These are then used to generate tickets, call an incident or trigger an automated remediation. By combining event management with Watchdog, Bits AI and workflow automations, Datadog now provides a full AI solution that helps teams automate remediation, proactively prevent outages and reduce the impact of incidents…

…We just announced in GA, the Event Management product, which is the main missing building block we had for AIOps platform.

Datadog’s Azure business is growing faster than Azure itself, and Datadog’s AI-part of Azure is growing than faster then the AI-part of Azure itself

The hyperscaler that is the most open by is — or transparent by in terms of numbers is Microsoft as they disclose how much of their growth comes from AI more specifically. And I will say that if you compare our business to theirs, the Azure part of our business is growing faster than Azure itself. And the AI-driven part of our Azure business itself is also growing faster than what you see on the on the overall Azure number. So we think we have similar exposure, and we track to the same trends broadly

Datadog’s AI exposure leans toward inferencing and applications in production a lot more than the training of models

I will say also on AI adoption that some of the revenue jumps you might see from the cloud providers might relate to supply of GPUs coming online and a lot of training clusters being provisioned. And those typically won’t generate a lot of new usage for us. We tend to be more correlative with the live applications, production applications and inference workloads that tend to follow after that, and that are more tied to all of these applications going into production. 

Datadog has products for monitoring what AI models are doing, but those products are not in general availability yet; management expects to have more announcements on these monitoring products in the near future; Datadog’s customers that are the most scaled on AI workloads are model providers, and they tend to have their own monitoring infrastructure for the quality of the models; the needs of the model providers for monitoring infrastructure are not representative of the needs of the bulk of the market, but there may still be overlaps in the future if the situation with the cloud hyperscalers is used as a guide

We have products for monitoring, not just the infrastructure, but what the LLMs are doing. Those products are still not in GA, so we’re working with a smaller number of design partners for that. As I think not only these products are maturing, but also the industry around us is maturing and more of these applications are getting into production. You should expect to hear more from us on that topic in the near future. The customers we have that are the most scaled on AI workloads are the model providers themselves, and they tend to have their own infrastructure for monitoring the quality of the models…

…On the tooling, I would say there’s a handful of players that have been building that tooling for a few years for — in a way that’s very specialized to what they do internally. They are not necessarily the representative of the bulk of the market. So in those situations, we’re always careful about overfitting products to a group that might not be the right target customer group in the end in the same way that building infrastructure monitoring for the cloud providers to use internally might not be an exact fit for what the rest of the world needs. That being said, I mean, look, we work a lot with those companies, and they have a number of needs that some of them they can meet internally and some of them, they don’t. And if I go back to the example of hyperscalers, we actually have teams at the hyperscalers that use us for application or infrastructure or logs internally, even though they’ve built a lot of that tooling themselves. So I think everything is possible in the long run. But our focus is really on the vast majority of the customer base that’s going to either use those API-based products or tune and run their own models.  

Datadog’s management is seeing a trend of AI-adopters starting with an API-accessible AI model to build applications, before offloading some of the workload to open-sourced AI models

We think there are good/bad weather in terms of what the adoption of AI is going to be from all the other companies, and we definitely see a trend where customers start with an API-driven or API-accessible model, build applications and then offload some of that application to other models that typically come from the open source and they might train, fine-tune themselves to get to a lower cost and lower time to respond.

Management is seeing a lot of interest in Datadog’s new AI-related products; management thinks its AI-related products are a joy to use

We see a lot of interest in the new products. These are new products so we just announced in GA, the Event Management product, which is the main missing building block we had for AIOps platform. And we also just released into GA, Bits for incident management. So there’s a lot of demand for it. The products are actually, I will say it, for Bits for incident management is a joy to use.

Etsy (NASDAQ: ETSY)

Etsy’s management believes the company’s product team is getting more efficient with machine learning

We had double-digit growth in the number of experiments per product engineer that utilize machine learning as well as in our annualized gross GMS from experiments. And the total number of experiments run per engineer increased 20%. Some of this progress can be directly tied to work we told you about last year to democratize ML. These metrics give me confidence that the bold moves to improve customer experience can build over time and play a key role to get Etsy growing again.

Etsy’s management thinks the application of AI is very useful for the company’s Gift Mode initiative

Large language models were really helpful for Gift Mode. So for example, there are 200 different persona in Gift Mode. And then within each persona, there are 3 to 5 different gift ideas and the ability to ask large language models, what are 200 examples of persona, and it wasn’t quite this simple, but it does give you a head start on that. If I’m a foodie who also loves to travel, what are 3 things I might buy on Etsy, 3 different ideas for gifts on Etsy, like, it does help to come up with a lot of ideas more quickly. The productivity gains, large language models are starting to help us with coding productivity as well. 

Etsy’s management finds the use of machine learning (ML) to be particularly useful in removing products that violate Etsy’s policies

We’re doing more than ever to suppress and remove listings that violate our policies. And advances in ML have been particularly powerful as enablers here. In the first quarter, we removed about 115% more listings for violating our handmade policy than in the prior year…

…For example, does this same item exist also on AliExpress. And we assume right now, if that item exists on AliExpress, we assume it’s mass produced and we take it down. You as a seller can appeal that, you can tell us how you made it yourself, and it still ended up on AliExpress. And by the way, that’s true sometimes. You can appeal that, but our default now is we take that down. And that’s just one example. Gen AI is actually going to be, I think, more and more helpful at understanding how much value did this particular seller truly add to the product.

Etsy’s management has used machine learning to improve the estimation of delivery time for products

In terms of shipping timeliness, I’m pleased to report that our initiative to tighten estimated delivery dates, which we believe are an important effort to improve buyer perceptions of our reliability as well as to grow GMS, are already paying off. Our fulfillment team recently launched a new machine learning model, which reduced our estimate of USPS transit times by greater than 1 day, resulting in a nearly tripling of the percentage of eligible orders for which Etsy is now able to show an estimated delivery date of 7 days or less.

Fiverr (NYSE: FVRR)

Fiverr’s management continues to see AI having a net positive impact on the company’s business; AI-related services saw 95% year-on-year growth in GMV on Fiverr’s platform, with chatbot development being especially popular; a hospitality company and an online learning platform are two examples of companies that have used Fiverr for AI chatbot development

AI continued to have a net positive impact on our business, as complex services continue to grow faster and represent a bigger portion of our business. Demand for AI-related services remained strong, as evidenced by 95% year-over-year growth in GMV from AI service categories. Chatbot development was especially popular this quarter as businesses look for ways to lean into GenAI technology to better engage with customers. For example, we have seen a hospitality company building a conversational tool for customers to manage bookings or an online learning platform creating a personalized learning menu and tutoring sessions for children. 

Fiverr has a pool of 10,000 AI experts and it is growing

With an over 10,000 and growing AI expert pool, Fiverr has become the destination for businesses to get help implementing GenAI and take their business to the next level.

Fiverr’s management is seeing very promising signals on Fiverr Neo, the company’s AI assistant for matching buyers with sellers; one-third of buyers who received seller recommendations from Neo sent a project brief to a seller; overall order conversion with Neo is nearly 3x that of the Fiverr marketplace average; management is excited about the potential of AI matching technology 

We have also seen very promising signals on Fiverr Neo, the AI matching assistant that we launched last year. Neo enables our buyers to have a more natural purchasing path by creating a conversational experience that leverages the catalog data and search algo. Answers and steps are provided based on buyers questions and the stage of the search. As a result, we saw that nearly one-third of the buyers who received seller recommendations from Neo ended up sending a project brief to the seller and the overall order conversion is nearly 3 times that of the marketplace average. This really gives us confidence and excitement in the potential we could unlock by investing in AI matching technology.

Fiverr’s product innovation pace had picked up in recent years; the latest set of product innovations will be focused on deepening trust and leveraging AI

Our product innovation pace picked up even more in recent years as the scale of our marketplace significantly expanded. This includes monetization products, such as Promoted Gigs and Seller Plus; AI innovations such as Logo Maker, AI Audition, to the latest ground-breaking Fiverr Neo; Business Solutions offerings, such as Project Partner and Fiverr Certified; and numerous products and features such as Fiverr Discover, Milestones and Subscriptions that empower our community to work better and smarter. We are always leading the curve of innovation that powers growth not only for us, but for the industry.

As our teams work towards our July product release, we are focusing on deepening trust and leveraging AI to reimagine every aspect of the customer journey. This includes improving our catalog and building new experiences to enable high-stakes, high-trust work to happen on Fiverr. We are strengthening our muscle in knowing our customers better in order to provide them with the better matching, better recommendations and better customer care, all of which leads to more trust for Fiverr as a platform. We are already seeing some of the benefits in unlocking wallet share and driving a mix shift towards complex services on Fiverr, and we are going to see more impact down the road

All the work that Fiverr facilitates happens on Fiverr, so management believes that the company has a lot of data (for perspective, in 2023, 38 million files were exchanged on Fiverr, and 2.5 million messages were sent daily between buyers and sellers) to leverage with generative AI to take the matching experience for buyers and sellers to a new level

Second, data and AI matching. Fiverr is unique in the sense that we are not just a platform that connects businesses with freelancers, the entire work actually happens on Fiverr. And that is really the secret sauce that enables us to do matching in such a simple, accurate and seamless way. With Generative AI, there’s incredible potential to take that experience to a whole new level. Just to give you some idea of the scale we operate. In 2023, over 38 million files were exchanged on our platform, and on average, 2.5 million messages were sent between buyers and sellers on a daily basis. We are experimenting with GenAI technology on how to unlock the potential of that massive data on Fiverr in order to enable buyers and sellers to have more information, search and browse in new ways, ask more complex questions, and ultimately, make better, more informed choices on Fiverr.

Fiverr’s management is seeing the presence of AI having a negative impact on the simple, low-value services on the company’s marketplace, but AI is overall a net-positive for Fiverr; management gave an example of how only simple language translation services are being impacted by AI, but the complex translation services are not

We mentioned in the previous earnings the fact that the negative impact that we’re seeing from AI is mostly around the very simple types of services. Those are normally services that would sell for $10, $15, which is — I mean, we are moving. I mean, the majority of contribution is coming from more complex services anyway. And as I said, we continue to see AI as a net positive. So it’s contributing more than the offsetting factors of simple products.It happens across several categories in several verticals, but there’s nothing specific to call out. Even if you look at the areas that you might think that AI would influence significantly like translation. But what you’re seeing is actually the very simple services around translation are being affected, the more complex types of services are not. I mean, if you would publish a book and then want to translate it into a different language that you don’t command, I would doubt that you would let AI translate it and go publish the outcome without actually verifying it.

Fiverr’s management is sure that many experts use AI as part of their workflow, but they do not rely on the AI blindly

I’m sure many experts actually use AI tools in their process of work, but they don’t rely on blindly letting AI run the work for them, but it is more of the modern tech that they use in order to amplify their creative process.

Mastercard (NYSE: MA)

Scam Protect is a new service launched by Mastercard’s management to protect users against cybercrime; Scam Protect combines Mastercard’s identity biometric AI and open banking capabilities

Cybercrime is a growing concern, last year alone, people in the United States lost over $12 billion to Internet scams. Scam Protect builds on the cybersecurity protections we have delivered for years, combines our identity biometric AI and open banking capabilities to identify and prevent scans before they occur. 

Mastercard is partnering with Verizon to design new AI tools to identify and block scammers

By combining Mastercard’s Identity Insights with Verizon’s robust network technologies, new AI power tools will be designed to more accurately identify and block scammers. 

Mastercard’s management has continued to enhance the company’s solutions with generative AI; Decision Intelligence Pro is a real-time transaction fraud solution for banks that is powered by generative AI to improve scoring and fraud detection by 20%; management sees tremendous opportunity with generative AI and has created a central role for AI

We continue to enhance our solutions with generative AI to deliver even more value, a world-leading real-time fraud solution, Decision Intelligence, has been helping banks score and safely approve billions of transactions, ensuring the safety of consumers and the entire payments networks for years. The next-generation technology, Decision Intelligence Pro is supercharged by generative AI to improve the overall score and boost fraud detection rates on average by 20%…

…We see tremendous opportunity on the AI side, particularly on the generative AI side, and we’ve created a central role for that. 

Meta Platforms (NASDAQ: META)

Meta is building a number of different AI services, including Meta AI (an AI assistant), creator AIs, business AIs, internal coding and development AIs, and hardware for AI interactions

We are building a number of different AI services from Meta AI, our AI assistant that you can ask any question across our apps and glasses, to creator AIs that help creators engage their communities and that fans can interact with, to business AIs that we think every business eventually on our platform will use to help customers buy things and get customer support, to internal coding and development AIs, to hardware like glasses for people to interact with AIs and a lot more.

Meta’s management released the company’s new version of Meta AI recently and it is powered by the company’s latest foundational model, Llama 3; management’s goal is for Meta AI to be the world’s leading AI service; tens of millions of people have tried Meta AI and the user feedback has been very positive; Meta AI is currently in English-speaking countries, but will be rolled out in more languages and countries in the coming months; management believes that the Llama3 version of Meta AI is the most intelligent AI assistant; Meta AI can be used within all of Meta’s major apps; besides being able to answer queries, Meta AI can also create animations as well as generate images while users are typing, which is a magical experience; Meta AI can also be used in Search within Meta’s apps, and Feed and Groups on Facebook

Last week, we had the major release of our new version of Meta AI that is now powered by our latest model, Llama 3. And our goal with Meta AI is to build the world’s leading AI service, both in quality and usage. The initial rollout of Meta AI is going well. Tens of millions of people have already tried it. The feedback is very positive. And when I first checked in with our teams, the majority of feedback we were getting was people asking us to release Meta AI for them wherever they are. So we’ve started launching Meta AI in some English speaking countries, and we’ll roll out in more languages and countries over the coming months…

…We believe that Meta AI with Llama 3 is now the most intelligent AI assistant that you can freely use. And now that we have the superior quality product, we’re making it easier for lots of people to use it within WhatsApp, Messenger, Instagram, and Facebook…

…In addition to answering more complex queries, a few other notable and unique features from this

release: Meta AI now creates animations from still images, and now generates high quality images so

fast that it can create and update them as you’re typing, which is pretty awesome. I’ve seen a lot of people commenting about this experience online and how they’ve never seen or experienced anything like it before…

…Along with using Meta AI within our chat surfaces, people will now be able to use Meta AI in Search within our apps, as well as Feed and Groups on Facebook. We expect these integrations will complement our social discovery strategy as our recommendation systems help people to discover and explore their interests while Meta AI enables them to dive deeper on topics they’re interested in. 

Meta’s foundational AI model, Llama3, has three versions with different number of parameters; management thinks the two smaller versions are both best-in-class for their scale; the 400+ billion parameter version of Llama3 is still undergoing training and is on track to be industry-leading; management thinks the Llama3 models will improve from further open source contributions

I’m very pleased with how Llama 3 has come together so far. The 8B and 70B parameter models that we released are best-in-class for their scale. The 400+B parameter model that we’re still training seems on track to be industry-leading on several benchmarks. And I expect that our models are just going to improve further from open source contributions. 

Meta’s management wants the company to invest significantly more in the coming years to build more advanced AI models and the largest scale AI services in the world, but the AI investments will come ahead of any meaningful revenue-generation from these new AI products

This leads me to believe that we should invest significantly more over the coming years to build even more advanced models and the largest scale AI services in the world. As we’re scaling capex and energy expenses for AI, we’ll continue focusing on operating the rest of our company efficiently. But realistically, even with shifting many of our existing resources to focus on AI, we’ll still grow our investment envelope meaningfully before we make much revenue from some of these new products…

… …We anticipate our full-year 2024 capital expenditures will be in the range of $35-40 billion, increased from our prior range of $30-37 billion as we continue to accelerate our infrastructure investments to support our AI roadmap. While we are not providing guidance for years beyond 2024, we expect capex will continue to increase next year as we invest aggressively to support our ambitious AI research and product development efforts.

Meta’s management thinks there are a few ways to build a massive AI business for Meta – these include business messaging, introducing ads and paid content in AI interactions, and selling access to powerful AI models and AI compute – in addition to the benefits to Meta’s current digital advertising business through the use of AI; management thinks business messaging is one of Meta’s nearer-term opportunities; management’s long-term vision for business messaging is to have AI agents that can accomplish goals rather than merely be a chatbot that replies to messages; management thinks that the capabilities of Meta’s business messaging AI technology will see massive improvements in as short as a year’s time

There are several ways to build a massive business here, including scaling business messaging, introducing ads or paid content into AI interactions, and enabling people to pay to use bigger AI models and access more compute. And on top of those, AI is already helping us improve app engagement which naturally leads to seeing more ads, and improving ads directly to deliver more value…

… The cost of engaging with people in messaging is still very high. But AI should bring that down just dramatically for businesses and creators. And I think that, that has the potential. That’s probably the — beyond just increasing engagement and increasing the quality of the ads, I think that, that’s probably one of the nearer-term opportunities, even though that will — it’s not like next quarter or the quarter after that scaling thing, but it’s — but that’s not like a 5-year opportunity either…

…I think that the next phase for a lot of these things are handling more complex tasks and becoming more like agents rather than just chat bots, right? So when I say chatbot, what I mean is if you send a message and it replies to your message, right? So it’s almost like almost a 1:1 correspondence. Whereas what an agent is going to do is you give it an intent or a goal, then it goes off and probably actually performs many queries on its own in the background in order to help accomplish your goal, whether that goal is researching something online or eventually finding the right thing that you’re looking to buy…  I think basically, the larger models and then the more advanced future versions that will be smaller as well are just going to enable much more interesting interactions like that. So I mean if you think about this, I mean, even some of the business use cases that we talked about, you don’t really just want like sales or customer support chatbot that can just respond to what you say. If you’re a business, you have a goal, right? You’re trying to support your customers well and you’re trying to position your products in a certain way and encourage people to buy certain things that map to their interests and would they be interested in? And that’s more of like a multiturn interaction, right?

So the type of business agent that you’re going to be able to enable with just a chatbot is going to be very naive compared to what we’re going to have in a year even, but beyond that, too, is just the reasoning and planning abilities if these things grow to be able to just help guide people through the business process of engaging with whatever your goals are as a creator of a business. So I think that that’s going to be extremely powerful. 

Meta’s AI recommendation system is currently delivering 30% of posts on the Facebook feed (up 2x over the last few years) and more than 50% of the content people see on Instagram (the first time this threshold is reached)

Right now, about 30% of the posts on Facebook feed are delivered by our AI recommendation system. That’s up 2x over the last couple of years. And for the first time ever, more than 50% of the content people see on Instagram is now AI recommended.

Revenue from two of Meta’s end-to-end AI-powered advertising tools, Advantage+ Shopping and Advantage+ App Campaigns, have more than doubled since last year; test results for the single-step automation feature of Advantage+ has resulted in a 28% decrease in cost per click or per objective for advertisers; Meta has significant runway to broaden adoption of the end-to-end automation features of Advantage+ and the company has enabled more conversion types

If you look at our two end-to-end AI-powered tools, Advantage+ Shopping and Advantage+ App Campaigns, revenue flowing through those has more than doubled since last year…

…So on the single-step automation, Advantage Plus audience, for example, has seen significant growth in adoption since we made it the default audience creation experience for most advertisers in Q4, and that enables advertisers to increase campaign performance by just using audience inputs as a suggestion rather than a hard constraint. And based on tests that we ran, campaigns using Advantage Plus audience targeting saw on average, a 28% decrease in cost per click or per objective compared to using our regular targeting.

On the end-to-end automation products like Advantage Plus shopping and Advantage Plus app campaigns, we’re also seeing very strong growth…  We think there’s still significant runway to broaden adoption, so we’re trying to enable more conversion types for Advantage Plus shopping. In Q1, we began expanding the list of conversions that businesses could optimize for. So previously, it only supported purchase events, and now we’ve added 10 additional conversion types. And we’re continuing to see strong adoption now across verticals.

Meta’s management continues to develop Meta’s own AI chips; Meta’s Training and Inference Accelerator chip is less expensive for Meta and has already been running some of Meta’s recommendation workloads

We’ll also keep making progress on building more of our own silicon. Our Meta Training and Inference Accelerator chip has successfully enabled us to run some of our recommendations-related workloads on this less expensive stack, and as this program matures over the coming years we plan to expand this to more of our workloads as well.

Meta’s management sees a market for a fashionable pair of AI glasses without holographic displays; management thinks that glasses are the ideal device for an AI assistant because the glasses can see what you see and hear what you hear; management recently launched Meta AI with Vision on its AI glasses; Meta’s AI glasses continue to do well and are sold out in many styles and colours

I used to think that AR glasses wouldn’t really be a mainstream product until we had full holographic displays — and I still think that will be awesome and is mature state of the product. But now it seems pretty clear that there’s also a meaningful market for fashionable AI glasses without a display. Glasses are the ideal device for an AI assistant because you can let them see what you see and hear what you hear, so they have full context on what’s going on around you as they help you with whatever you’re trying to do. Our launch this week of Meta AI with Vision on the glasses is a good example where you can now ask questions about things you’re looking at…

…The Ray-Ban Meta glasses that we built with Essilor Luxottica continue to do well and are sold out in many styles and colors, so we’re working to make more and release additional styles as quickly as we can.

Meta’s management is improving the monetisation efficiency of the company’s products partly by using larger AI models in its new ads ranking architecture, Meta Lattice (which was rolled out last year) in place of smaller models, as well as using AI to provide more automation – ranging from point-automation to end-to-end automation – for advertisers through its Advantage+ portfolio; Meta Lattice drove improved ad performance over the course of 2023 when it was deployed across Facebook and Instagram

The second part of improving monetization efficiency is enhancing marketing performance. Similar to our work with organic recommendations, AI is playing an increasing role in these efforts. First, we are making ongoing ads modeling improvements that are delivering better performance for advertisers. One example is our new ads ranking architecture, Meta Lattice, which we began rolling out more broadly last year. This new architecture allows us to run significantly larger models that generalize learnings across objectives and surfaces in place of numerous, smaller ads models that have historically been optimized for individual objectives and surfaces. This is not only leading to increased efficiency as we operate fewer models, but also improving ad performance. Another way we’re leveraging AI is to provide increased automation for advertisers. Through our Advantage+ portfolio, advertisers can automate one step of the campaign set up process – such as selecting which ad creative to show – or automate their campaign completely using our end-to-end automation tools, Advantage+ Shopping and Advantage+ App ads. We’re seeing growing use of these solutions, and we expect to drive further adoption over the course of the year while applying what we learn to our broader ads investments…

…We’ve talked a little bit about the new model architecture at Meta Lattice that we deployed last year that consolidates smaller and more specialized models into larger models that can better learn what characteristics improve ad performance across multiple services, like Feed and Reels and multiple types of ads and objectives at the same time. And that’s driven improved ad performance over the course of 2023 as we deployed it across Facebook and Instagram to support multiple objectives.

Meta’s recommendation products historically each had their own AI models, and a new model architecture to power multiple recommendation products was being developed recently; the new model architecture was tested last year on Facebook Reels and generated 8%-10% increases in watch time; the new model architecture has been extended beyond Reels and management is hopeful that the new architecture will unlock better video recommendations over time

Historically, each of our recommendation products, including Reels, in-feed recommendations, et cetera, has had their own AI model. And recently, we’ve been developing a new model architecture with the aim for it to power multiple recommendations products. We started partially validating this model last year by using it to power Facebook Reels. And we saw meaningful performance gains, 8% to 10% increases in watch time as a result of deploying this. This year, we’re actually planning to extend the singular model architecture to recommend content across not just Facebook Reels, but also Facebook’s video tab as well. So while it’s still too early to share specific results, we’re optimistic that the new model architecture will unlock increasingly relevant video recommendations over time. And if it’s successful, we’ll explore using it to power other recommendations.

Meta’s management is seeing adoption of Meta’s generative AI (GenAI) ad creative features across verticals and different advertiser sizes; some of these features are enjoying outsized adoption; Meta expects improvements to its underlying foundational AI models to improve the output quality of its GenAI ad creative features

The more near-term version is around the GenAI ad creative features that we have put into our ads creation tools. And it’s early, but we’re seeing adoption of these features across verticals and different advertiser sizes. In particular, we’ve seen outsized adoption of image expansion with small businesses, and this will remain a big area of focus for us in 2024, and I expect that improvements to our underlying foundation models will enhance the quality of the outputs that are generated and support new features on the road map. But right now, we have features supporting text variations, image expansion and background generation, and we’re continuing to work to make those more performance for advertisers to create more personalized ads at scale.

In early tests of using business AIs for business messaging, Meta’s management is receiving positive feedback from users

The longer-term piece here is around business AIs. We have been testing the ability for businesses to set up AIs for business messaging that represent them in chats with customers starting by supporting shopping use cases such as responding to people asking for more information on a product or its availability. So this is very, very early. We’ve been testing this with a handful of businesses on Messenger and WhatsApp, and we’re hearing good feedback with businesses saying that the AIs have saved them significant time while customer — consumers noted more timely response times. And we’re also learning a lot from these tests to make these AIs more performant over time as well.

Meta’s management has gotten more optimistic and ambitious on AI compared to just 3 months ago because of the company’s work with Llama3 and Meta AI

[Question] Can you just talk about what’s changed most in your view in the business and the opportunity now versus 3 months ago? 

[Answer]  I think we’ve gotten more optimistic and ambitious on AI. So previously, I think that our work in this — I mean when you were looking at last year, when we released Llama 2, we were very excited about the model and thought that, that was going to be the basis to be able to build a number of things that were valuable that integrated into our social products. But now I think we’re in a pretty different place. So with the latest models, we’re not just building good AI models that are going to be capable of building some new good social and commerce products. I actually think we’re in a place where we’ve shown that we can build leading models and be the leading AI company in the world. And that opens up a lot of additional opportunities beyond just ones that are the most obvious ones for us. So that’s — this is what I was trying to refer to in my opening remarks where I just view the success that we’ve seen with the way that Lama 3 and Meta AI have come together as a real validation technically that we have the talent, the data and the ability to scale infrastructure to do leading work here.

Meta’s AI capex can be categorised into 2 buckets, with one being core AI work that has a very ROI-driven (return on investment driven) approach and which still generates very strong returns, and the other being generative AI and other advanced research work that has tremendous potential but has yet to produce returns; Meta’s AI capex for the 2 buckets are in capacity that is fungible

We’ve broadly categorized our AI investments into 2 buckets. I think of them as sort of core AI work and then strategic bets, which would include Gen AI and the advanced research efforts to support that. And those are just really at different stages as it relates to being able to measure the return and drive revenue for our business.

So with our core AI work, we continue to have a very ROI-driven approach to investment, and we’re still seeing strong returns as improvements to both engagement and ad performance have translated into revenue gains.

Now the second area, strategic bets, is where we are much earlier. Mark has talked about the potential that we believe we have to create significant value for our business in a number of areas, including opportunities to build businesses that don’t exist on us today. But we’ll need to invest ahead of that opportunity to develop more advanced models and to grow the usage of our products before they drive meaningful revenue. So while there is tremendous long-term potential, we’re just much earlier on the return curve than with our core AI work.

What I’ll say though is we’re also building our systems in a way that gives us fungibility in how we use our capacity, so we can flex it across different use cases as we identify what are the best opportunities to put that infrastructure toward.

Meta is already shifting a lot of resources from other parts of the company into its AI efforts

I would say broadly, we actually are doing that in a lot of places in terms of shifting resources from other areas, whether it’s compute resources or different things in order to advance the AI efforts. 

Meta has partnered with Google and Bing for Meta AI’s search citations, but management has no intention to build a search ads business

[Question] You partnered with Google and Bing for Meta AI organic search citations. So I guess stepping back, do you think that Meta AI longer term could bring in search advertising dollars at some point?

[Answer] On the Google and Microsoft partnerships, yes, I mean we work with them to have real-time information in Meta AI. It’s useful. I think it’s pretty different from search. We’re not working on search ads or anything like that. I think this will end up being a pretty different business.

Microsoft (NASDAQ: MSFT)

Azure took market share again in 2024 Q1; Microsoft’s management thinks that (1) Azure offers the most diverse selection of AI accelerators, including those from Nvidia, AMD, and Microsoft’s own custom chips, (2) Azure offers the best selection of foundational AI models, including LLMs and SLMs (small language models), and (3) Azure’s Models as a Service offering makes it easy for developers to work with LLMs and SLMs without having to worry about technical infrastructure; >65% of Fortune 500 use Azure OpenAI service; hundreds of paid customers are using Azure’s Models as a Service to access third-party AI models including those from Cohere, Meta, and Mistral; Azure grew revenue by 31% in 2024 Q1 (was 30% in 2023 Q4), with 7 points of growth from AI services (was 6 points in 2023 Q4); Azure’s non-AI consumption business also saw broad greater-than-expected demand 

Azure again took share as customers use our platforms and tools to build their own AI solutions. We offer the most diverse selection of AI accelerators, including the latest from NVIDIA, AMD as well as our own first-party silicon…

…More than 65% of the Fortune 500 now use Azure OpenAI service. We also continue to innovate and partner broadly to bring customers the best selection of frontier models and open source models, LLMs and SLMs…

…Our Models as a Service offering makes it easy for developers to use LLMs and SLMs without having to manage any underlying infrastructure. Hundreds of paid customers from Accenture and EY to Schneider Electric are using it to take advantage of API access to third-party models including, as of this quarter, the latest from Cohere, Meta and Mistral…

… Azure and other cloud services revenue grew 31% ahead of expectations, while our AI services contributed 7 points of growth as expected. In the non-AI portion of our consumption business, we saw greater-than-expected demand broadly across industries and customer segments as well as some benefit from a greater-than-expected mix of contracts with higher in-period recognition. 

Microsoft’s management continues to build on the company’s partnership with OpenAI for AI work

Our AI innovation continues to build on our strategic partnership with OpenAI. 

Microsoft’s management thinks that Phi-3, announced by Microsoft recently, is the most capable and cost-effective SLM and it’s being trialed by a number of companies

With Phi-3, which we announced earlier this week, we offer the most capable and cost-effective SLM available. It’s already being trialed by companies like CallMiner, LTIMindtree, PwC and TCS.

Azure AI customers are growing and spending more with Microsoft; over half of Azure AI customers use Microsoft’s data and analytics tools and they are building applications with deep integration between these tools and Azure AI

All up, the number of Azure AI customers continues to grow and average spend continues to increase…

… Over half of our Azure AI customers also use our data and analytics tools. Customers are building intelligent applications running on Azure, PostgreSQL and Cosmos DB with deep integrations with Azure AI. TomTom is a great example. They’ve used Cosmos DB along with Azure OpenAI service to build their own immersive in-car infotainment system. 

GitHub Copilot now has 1.8 million paid subscribers, up 35% sequentially; even established enterprises are using GitHub Copilot; >90% of Fortune 100 companies are GitHub customers; GitHub’s revenue was up 45% year-on-year

GitHub Copilot is bending the productivity curve for developers. We now have 1.8 million paid subscribers with growth accelerating to over 35% quarter-over-quarter and continues to see increased adoption from businesses in every industry, including Itau, Lufthansa Systems, Nokia, Pinterest and Volvo Cars. Copilot is driving growth across the broader GitHub platform, too. AT&T, Citigroup and Honeywell all increased their overall GitHub usage after seeing productivity and code quality increases with Copilot. All up, more than 90% of the Fortune 100 are now GitHub customers, and revenue accelerated over 45% year-over-year.

Microsoft has new AI-powered features within its low-code and no-code tools for building applications; 30,000 organisations – up 175% sequentially – across all industries have used Copilot Studio to customise or build their own copilot; Cineplex used Copilot Studio to build a copilot for customer service agents to significantly reduce the time needed to handle queries; Copilot Studio can be really useful for enterprises to ground their AIs with enterprise data, and people are really excited about it

Anyone can be a developer with new AI-powered features across our low-code, no-code tools, which makes it easier to build an app, automate workflow or create a Copilot using natural language. 30,000 organizations across every industry have used Copilot Studio to customize Copilot for Microsoft 365 or build their own, up 175% quarter-over-quarter. Cineplex, for example, built a Copilot for customer service agents, reducing query handling time from as much as 15 minutes to 30 seconds…

…Copilot Studio is really off to the races in terms of the product that most people are excited because one of the things in the enterprise is you want to ground your copilot with the enterprise data, which is in all of these SaaS applications, and Copilot Studio is the tool to use there to make that happen.

More than 330,000 organisations, including half of the Fortune 100, have used AI-features within Microsoft’s Power Platform

All up, over 330,000 organizations, including over half of Fortune 100, have used AI-powered capabilities in Power Platform, and Power Apps now has over 25 million monthly active users, up over 40% year-over-year.

In 2024 Q1, Microsoft’s management made Copilot available to all organisations; nearly 60% of Fortune 500 are using Copilot; many large companies have purchased more than 10,000 Copilot seats each; management is seeing higher usage of Copilot from early adopters, including a 50% jump in Copilot-assisted interactions per user in Teams; Microsoft has added more than 150 Copilot capabilities since the start of the year, including Copilot for Service, Copilot for Sales, Copilot for Finance, and Copilot for Security

This quarter, we made Copilot available to organizations of all types and sizes from enterprises to small businesses. Nearly 60% of the Fortune 500 now use Copilot, and we have seen accelerated adoption across industries and geographies with companies like Amgen, BP, Cognizant, Koch Industries, Moody’s, Novo Nordisk, NVIDIA and Tech Mahindra purchasing over 10,000 seats. We’re also seeing increased usage intensity from early adopters, including a nearly 50% increase in the number of Copilot-assisted interactions per user in Teams, bridging group activity with business process workflows and enterprise knowledge…

…We’re accelerating our innovation, adding over 150 Copilot capabilities since the start of the year…

… This quarter, we made our Copilot for Service and Copilot for Sales broadly available, helping customer service agents and sellers at companies like Land O’Lakes, Northern Trust, Rockwell Automation and Toyota Group generate role-specific insights and recommendations from across Dynamics 365 and Microsoft 365 as well as third-party platforms like Salesforce, ServiceNow and Zendesk. And with our Copilot for Finance, we are drawing context from Dynamics as well as ERP systems like SAP to reduce labor-intensive processes like collections and contract and invoice capture for companies like Dentsu and IDC…

…A great example is Copilot for Security, which we made generally available earlier this month, bringing together LLMs with domain-specific skills informed by our threat intelligence and 78 trillion daily security signals to provide security teams with actionable insights.

Microsoft’s management is seeing ISVs (independent software vendors) build their own Copilot integrations, with Adobe being an example

ISVs are also building their own Copilot integrations. For example, new integrations between Adobe Experience Cloud and Copilot will help marketeers access campaign insights in the flow of their work. 

Copilot in Windows is now available on 225 million PCs, up 2x sequentially; Microsoft’s largest PC partners have announced AI PCs in recent months; management recently introduced new Surface devices that comes with NPUs (neural processing units) that can power on-device AI experiences; management thinks that the presence of Copilot can help Microsoft create a new device-category for AI

When it comes to devices, Copilot in Windows is now available on nearly 225 million Windows 10 and Windows 11 PCs, up 2x quarter-over-quarter. With Copilot, we have an opportunity to create an entirely new category of devices purpose built for this new generation of AI. All of our largest OEM partners have announced AI PCs in recent months. And this quarter, we introduced new Surface devices, which includes integrated NPUs to power on device AI experiences like auto framing and live captions. And there’s much more to come. In just a few weeks, we’ll hold a special event to talk about our AI vision across Windows and devices.

More than 200 healthcare organisations are using Microsoft’s DAX Copilot

In health care, DAX Copilot is being used by more than 200 health care organizations, including Providence, Stanford Health care and WellSpan Health. 

Established auto manufacturers are using Microsoft’s AI solutions to improve their factory operations

And in manufacturing, this week at Hannover Messe, customers like BMW, Siemens and Volvo Penta shared how they’re using our cloud and AI solutions to transform factory operations.

LinkedIn AI-assisted messages have a 40% higher acceptance rate and are accepted >10% faster by job seekers; LinkedIn’s AI-powered collaborative articles now have more than 12 million contributions and helped engagement on LinkedIn reach a new record in 2024 Q1; LinkedIn Premium’s revenue was up 29% year-on-year in 2024 Q1, with AI features helping to produce the growth

Features like LinkedIn AI-assisted messages are seeing a 40% higher acceptance rate and accepted over 10% faster by job seekers saving hirers time and making it easier to connect them to candidates. Our AI-powered collaborative articles, which has reached over 12 million contributions are helping increase engagement on the platform, which reached a new record this quarter. New AI features are also helping accelerate LinkedIn Premium growth with revenue up 29% year-over-year. 

Microsoft’s management expects capex to increase materially sequentially in 2024 Q2 (FY2024 Q4) because of cloud and AI infrastructure investments; management sees near-term AI demand as being higher than available capacity; capex in FY2025 is expected to be higher than in FY2024, but this will be driven ultimately by the amount of AI inference demand; operating margin in FY2025 is expected to be down by only 1 point compared to FY2024

We expect capital expenditures to increase materially on a sequential basis driven by cloud and AI infrastructure investments. As a reminder, there can be normal quarterly spend variability in the timing of our cloud infrastructure build-outs and the timing of finance leases. We continue to bring capacity online as we scale our AI investments with growing demand. Currently, near-term AI demand is a bit higher than our available capacity…

…In FY ’25, that focus on execution should again lead to double-digit revenue and operating income growth. To scale to meet the growing demand signal for our cloud and AI products, we expect FY ’25 capital expenditures to be higher than FY ’24. These expenditures over the course of the next year are dependent on demand signals and adoption of our services, so we will manage that signal through the year. We will also continue to prioritize operating leverage. And therefore, we expect FY ’25 operating margins to be down only about 1 point year-over-year, even with our significant cloud and AI investments as well as a full year of impact from the Activision acquisition…

… Then, Amy referenced what we also do on the inference side, which is, one, we first innovate and build products. And of course, we have an infrastructure business that’s also dependent on a lot of ISVs building products that run on our infrastructure. And it’s all going to be demand driven. In other words, we track very closely what’s happening with inference demand, and that’s something that we will manage, as Amy said in her remarks, very, very closely.

Microsoft’s management expects Azure to grow revenue by 30%-31% in constant currency, similar to stronger-than-expected 2024 Q1 results, driven by AI

For Intelligent Cloud, we expect revenue to grow between 19% and 20% in constant currency or USD 28.4 billion to USD 28.7 billion. Revenue will continue to be driven by Azure, which, as a reminder, can have quarterly variability primarily from our per user business and in-period revenue recognition depending on the mix of contracts. In Azure, we expect Q4 revenue growth to be 30% to 31% in constant currency or similar to our stronger-than-expected Q3 results. Growth will be driven by our Azure consumption business and continued contribution from AI with some impact from the AI capacity availability noted earlier.

Management’s AI-related capital expenditure plans for Microsoft has two layers to it, namely, training and inference; for training, management wants Microsoft to have capacity to train large foundation models and stay a leader in that area; for inference, management is watching inference demand

[Question] It looks like Microsoft is on track to ramp CapEx over 50% year-on-year this year to over $50 billion. And there’s media speculation of more spending ahead with some reports talking about like $100 billion data center. So obviously, investments are coming well ahead of the revenue contribution, but what I was hoping for is that you could give us some color on how you as the management team try to quantify the potential opportunities that underlie these investments because they are getting very big. 

[Answer]  At a high level, the way we, as a management team, talk about it is there are 2 sides to this, right? There is training and there’s inference. What — given that we want to be a leader in this big generational shift and paradigm shift in technology, that’s on the training side. We want to be able to allocate the capital required to essentially be training these large foundation models and stay on the leadership position there. And we’ve done that successfully all the way today, and you’ve seen it flow through our P&L, and you can continue to see that going forward. Then, Amy referenced what we also do on the inference side, which is, one, we first innovate and build products. And of course, we have an infrastructure business that’s also dependent on a lot of ISVs building products that run on our infrastructure. And it’s all going to be demand driven. In other words, we track very closely what’s happening with inference demand, and that’s something that we will manage, as Amy said in her remarks, very, very closely.

Microsoft’s management feels good about demand for Azure, because (1) they think Azure is a market-share taker since it has become the go-to choice for anybody who is working on an AI project, (2) they are seeing that AI projects on Azure do not stop with just calling AI models and there are many other cloud computing services in Azure that are required, (3), there’s migration to Azure, and (4) the optimisation cycle from the recent past has given more budget for people to start new workloads

[Question] How would you characterize the demand environment? On one hand, you have bookings in Azure both accelerating year-over-year in the quarter, but we’re seeing a lot of future concern, hesitation from other vendors we all cover. So I think everyone would love to get your sense of budget health for customers this year.

[Answer] On the Azure side, which I think is what you specifically asked, we feel very, very good about the — we’re fundamentally a share taker there because if you look at it from our perspective, at this point, Azure has become a port of call for pretty much anybody who is doing an AI project. And so that’s sort of been a significant help for us in terms of acquiring even new customers…

…The second thing that we’re also seeing is AI just doesn’t sit on its own. So AI projects obviously start with calls to AI models, but they also use a vector database. In fact, Azure Search, which is really used by even ChatGPT, is one of the fastest growing services for us. We have Fabric integration to Azure AI and so — Cosmos DB integration. So the data tier, the dev tools is another place where we are seeing great traction. So we are seeing adjacent services in Azure that get attached to AI…

… lastly, I would say, migration to Azure as well. So this is not just all an AI story. 

We are also looking at customers — I mean, this is something that we have talked about in the past, which is there’s always an optimization cycle. But there’s also — as people optimize, they spend money on new project starts, which will grow and then they’ll optimize. So it’s a continuous side of it. So these are the 3 trends that are playing out on Azure in terms of what at least we see on demand side.

Microsoft’s management thinks that a good place to watch for the level of maturation for AI will be what’s happening in terms of standard issues for software teams; they are seeing Copilots increasingly becoming “standard issue” for software teams; they think companies will need to undergo a cultural shift to fully embrace AI tools and it will take some time, but the rate of adoption of Copilot is also faster than anything they have seen in the past

[Question] We’re seeing companies shifting their IT spending to invest in and learn about AI rather than receiving additional budgets for AI. At some point for AI to be transformative, as everyone expects, it needs to be accretive to spending. Satya, when do you believe AI will hit the maturity level?

[Answer] A good place to start is to watch what’s happening in terms of standard issues for software teams, right? I mean if you think about it, they bought tools in the past. Now you basically buy tools plus Copilot, right? So you could even say that this is characterized as perhaps shift of what is OpEx dollars into effectively tool spend because it gives operating leverage to all of the OpEx dollars you’re spending today, right? That’s really a good example of, I think, what’s going to happen across the board. We see that in customer service. We see that in sales. We see that in marketing, anywhere there’s operations…

…one of the interesting rate limiters is culture change inside of organizations. When I say culture change, that means process change…  That requires not just technology but in fact, companies to go do the hard work of culturally changing how they adopt technology to drive that operating leverage. And this is where we’re going to see firm-level performance differences…

…And so yes, it will take time to — for it to percolate through the economy. But this is faster diffusion, faster rate of adoption than anything we have seen in the past. As evidenced even by Copilot, right, it’s faster than any suite we have sold in the past.

Netflix (NASDAQ: NFLX)

Netflix has been working with machine learning (ML) for almost two decades, with ML being foundational for the company’s recommendation systems; management thinks that generative AI can be used to help creators improve their story-telling, and there will always be a place for creators

[Question]  What is the opportunity for Netflix to leverage generative AI technology in the near and long term? What do you think great storytellers should be focused on as this technology continues to emerge quickly? 

[Answer] Worth noting, I think, that we’ve been leveraging advanced technologies like ML for almost 2 decades. These technologies are the foundation for our recommendation systems that help us find these largest audiences for our titles and deliver the most satisfaction for members. So we’re excited to continue to involve and improve those systems as new technologies emerge and are developed.

And we also think we’re well positioned to be in the vanguard of adoption and application of those new approaches from our just general capabilities that we’ve developed and how we’ve already developed systems that do all these things.

We also think that we have the opportunity to develop and deliver new tools to creators to allow them to tell their stories in even more compelling ways. That’s great for them, it’s great for the stories, and it’s great for our members. 

And what should storytellers be focused on? I think storytellers should be focused on great storytelling. It is incredibly hard and incredibly complex to deliver thrilling stories through film, through series, through games. And storytellers have a unique and critical role in making that happen, and we don’t see that changing.

Nvidia (NASDAQ: NVDA)

Nvidia’s Data Center revenue had incredibly strong growth in 2024 Q1, driven by demand for the Hopper GPU computing platform; compute revenue was up by 5x while networking revenue was up by 3x

Data Center revenue of $22.6 billion was a record, up 23% sequentially and up 427% year-on-year, driven by continued strong demand for the NVIDIA Hopper GPU computing platform. Compute revenue grew more than 5x and networking revenue more than 3x from last year.

Nvidia’s management thinks that cloud providers are getting a 5x return on spending on Nvidia’s AI products over 4 years; management also thinks that cloud providers serving LLMs (large language models) via APIs (application programming interfaces) can earn $7 in revenue for every $1 spent on Nvidia’s H200 servers through running inference 

Training and inferencing AI on NVIDIA CUDA is driving meaningful acceleration in cloud rental revenue growth, delivering an immediate and strong return on cloud providers’ investment. For every $1 spent on NVIDIA AI infrastructure, cloud providers have an opportunity to earn $5 in GPU instant hosting revenue over 4 years…

…H200 nearly doubles the inference performance of H100, delivering significant value for production deployments. For example, using Llama 3 with 700 billion parameters, a single NVIDIA HGX H200 server can deliver 24,000 tokens per second, supporting more than 2,400 users at the same time. That means for every $1 spent on NVIDIA HGX H200 servers at current prices per token, an API provider serving Llama 3 tokens can generate $7 in revenue over 4 years.

Nvidia’s management sees Nvidia GPUs as offering the best time-to-train AI models, the lowest cost to train AI models, and the lowest cost to run inference on AI models

For cloud rental customers, NVIDIA GPUs offer the best time-to-train models, the lowest cost to train models and the lowest cost to inference large language models.

Leading LLM (large language model) providers are building on Nvidia’s AI infrastructure in the cloud

Leading LLM companies such as OpenAI, Adept, Anthropic, Character.ai, Cohere, Databricks, DeepMind, Meta, Mistral, XAi, and many others are building on NVIDIA AI in the cloud.

Tesla is using Nvidia’s GPUs for its FSD (Full Self Driving) version 12 software for AI-powered autonomous driving; Nvidia’s management sees automotive as the largest enterprise vertical within its Data Center business this year

We supported Tesla’s expansion of their training AI cluster to 35,000 H100 GPUs. Their use of NVIDIA AI infrastructure paved the way for the breakthrough performance of FSD version 12, their latest autonomous driving software based on Vision. NVIDIA Transformers, while consuming significantly more computing, are enabling dramatically better autonomous driving capabilities and propelling significant growth for NVIDIA AI infrastructure across the automotive industry. We expect automotive to be our largest enterprise vertical within Data Center this year, driving a multibillion revenue opportunity across on-prem and cloud consumption.

Meta Platform’s Llama3 LLM was trained on a large cluster of Nvidia GPUs

A big highlight this quarter was Meta’s announcement of Llama 3, their latest large language model, which was trained on a cluster of 24,000 H100 GPUs. Llama 3 powers Meta AI, a new AI assistant available on Facebook, Instagram, WhatsApp, and Messenger. Llama 3 is openly available and has kickstarted a wave of AI development across industries.

Nvidia’s management sees inferencing of AI models growing as generative AI makes its way into more consumer internet applications

As generative AI makes its way into more consumer Internet applications, we expect to see continued growth opportunities as inference scales both with model complexity as well as with the number of users and number of queries per user, driving much more demand for AI compute.

Nvidia’s management sees inferencing accounting for 40% of Data Center revenue over the last 4 quarters

In our trailing 4 quarters, we estimate that inference drove about 40% of our Data Center revenue. Both training and inference are growing significantly.

Nvidia’s management is seeing companies build AI factories (large clusters of AI chips); Nvidia worked with more than 100 customers in 2024 Q1 to build AI factories that range in size from hundreds to tens of thousands of GPUs

Large clusters like the ones built by Meta and Tesla are examples of the essential infrastructure for AI production, what we refer to as AI factories. These next-generation data centers host advanced full-stack accelerated computing platforms where the data comes in and intelligence comes out.  In Q1, we worked with over 100 customers building AI factories ranging in size from hundreds to tens of thousands of GPUs, with some reaching 100,000 GPUs.

Nvidia’s management is seeing growing demand from nations for AI infrastructure and they see revenue from sovereign AI reaching high single-digit billions in 2024

From a geographic perspective, Data Center revenue continues to diversify as countries around the world invest in sovereign AI. Sovereign AI refers to a nation’s capabilities to produce artificial intelligence using its own infrastructure, data, workforce, and business networks. Nations are building up domestic computing capacity through various models. Some are procuring and operating sovereign AI clouds in collaboration with state-owned telecommunication providers or utilities. Others are sponsoring local cloud partners to provide a shared AI computing platform for public and private sector use. For example, Japan plans to invest more than $740 million in key digital infrastructure providers, including KDDI, Sakura Internet, and SoftBank to build out the nation’s sovereign AI infrastructure. France-based Scaleway, a subsidiary of the Iliad Group, is building Europe’s most powerful cloud native AI supercomputer. In Italy, Swisscom Group will build the nation’s first and most powerful NVIDIA DGX-powered supercomputer to develop the first LLM natively trained in the Italian language. And in Singapore, the National Supercomputer Centre is getting upgraded with NVIDIA Hopper GPUs, while Singtel is building NVIDIA’s accelerated AI factories across Southeast Asia…

…From nothing the previous year, we believe sovereign AI revenue can approach the high single-digit billions this year.

Nvidia’s revenue in China is down significantly in 2024 Q1 because of export restrictions for leading AI chips; management expects to see strong competitive forces in China going forward

We ramped new products designed specifically for China that don’t require export control license. Our Data Center revenue in China is down significantly from the level prior to the imposition of the new export control restrictions in October. We expect the market in China to remain very competitive going forward.

Because of improvements in CUDA algorithms, Nvidia’s management has been able to drive a 3x improvement in LLM inference speed on the H100 chips, which translates to a 3x cost reduction when serving AI models

Thanks to CUDA algorithm innovations, we’ve been able to accelerate LLM inference on H100 by up to 3x, which can translate to a 3x cost reduction for serving popular models like Llama 3.

Nvidia’s management sees the demand for the company’s latest AI chips to well exceed supply into 2025

We are working to bring up our system and cloud partners for global availability later this year. Demand for H200 and Blackwell is well ahead of supply, and we expect demand may exceed supply well into next year.

Nvidia’s strong networking growth in 2024 Q1 was driven by Infiniband

Strong networking year-on-year growth was driven by InfiniBand. We experienced a modest sequential decline, which was largely due to the timing of supply, with demand well ahead of what we were able to ship. We expect networking to return to sequential growth in Q2.

Nvidia’s management has started shipping its own Ethernet solution for AI networking called Spectrum-X Ethernet; management believes that Spectrum-X is optimised for AI from the ground-up, and delivers 1.6x higher networking performance for AI workloads compared with traditional ethernet; Spectrum-X is already ramping with multiple customers, including in a GPU cluster with 100,000 GPUs; Spectrum-X opens a new AI networking market for Nvidia and management thinks it can be a multi-billion product within a year; management is going all-in on Ethernet for AI networking, but they still see Infiniband as the superior solution; Infiniband started as a computing fabric and became a network, whereas Ethernet was a network that is becoming a computing fabric

In the first quarter, we started shipping our new Spectrum-X Ethernet networking solution optimized for AI from the ground up. It includes our Spectrum-4 switch, BlueField-3 DPU, and new software technologies to overcome the challenges of AI on Ethernet to deliver 1.6x higher networking performance for AI processing compared with traditional Ethernet. Spectrum-X is ramping in volume with multiple customers, including a massive 100,000 GPU cluster. Spectrum-X opens a brand-new market to NVIDIA networking and enables Ethernet-only data centers to accommodate large-scale AI. We expect Spectrum-X to jump to a multibillion-dollar product line within a year…

…But we’re all in on Ethernet, and we have a really exciting road map coming for Ethernet. We have a rich ecosystem of partners. Dell announced that they’re taking Spectrum-X to market. We have a rich ecosystem of customers and partners who are going to announce taking our entire AI factory architecture to market.

And so for companies that want the ultimate performance, we have InfiniBand computing fabric. InfiniBand is a computing fabric, Ethernet to network. And InfiniBand, over the years, started out as a computing fabric, became a better and better network. Ethernet is a network and with Spectrum-X, we’re going to make it a much better computing fabric. And we’re committed, fully committed, to all 3 links, NVLink computing fabric for single computing domain, to InfiniBand computing fabric, to Ethernet networking computing fabric. And so we’re going to take all 3 of them forward at a very fast clip. 

Nvidia’s latest AI chip-platform, Blackwell, delivers 4x faster training speeds, 30x faster inference speeds, and 25x lower total cost of ownership, compared to the H100 chip and enables real-time generative AI on trillion-parameter LLMs; the Blackwell platform includes Nvidia’s Inifiniband and Ethernet switches; management has built Blackwell to be compatible with all kinds of data centers; the earliest deployers of Blackwell include Amazonn, Google, Meta, and Microsoft; Nvidia’s management is on a 1-year development rhythm with the Blackwell platform-family, so there will be a new version of Blackwell in the next 12 months

At GTC in March, we launched our next-generation AI factory platform, Blackwell. The Blackwell GPU architecture delivers up to 4x faster training and 30x faster inference than the H100 and enables real-time generative AI on trillion-parameter large language models. Blackwell is a giant leap with up to 25x lower TCO and energy consumption than Hopper. The Blackwell platform includes the fifth-generation NVLink with a multi-GPU spine and new InfiniBand and Ethernet switches, the X800 series designed for a trillion-parameter scale AI. Blackwell is designed to support data centers universally, from hyperscale to enterprise, training to inference, x86 to Grace CPUs, Ethernet to InfiniBand networking, and air cooling to liquid cooling. Blackwell will be available in over 100 OEM and ODM systems at launch, more than double the number of Hoppers launched and representing every major computer maker in the world… 

…Blackwell time-to-market customers include Amazon, Google, Meta, Microsoft, OpenAI, Oracle, Tesla, and XAi…

…I can announce that after Blackwell, there’s another chip. And we are on a 1-year rhythm.

Nvidia’s management has introduced AI software called Nvidia Inference Microservices that allow developers to quickly build and deploy generative AI applications across a broad range of use cases including text, speech, imaging, vision, robotics, genomics, and digital biology

We announced a new software product with the introduction of NVIDIA Inference Microservices, or NIM. NIM provides secure and performance-optimized containers powered by NVIDIA CUDA acceleration in network computing and inference software, including Triton and PrintServer and TensorRT-LLM with industry-standard APIs for a broad range of use cases, including large language models for text, speech, imaging, vision, robotics, genomics, and digital biology. They enable developers to quickly build and deploy generative AI applications using leading models from NVIDIA, AI21, Adept, Cohere, Getty Images, and Shutterstock, and open models from Google, Hugging Face, Meta, Microsoft, Mistral AI, Snowflake and Stability AI. NIMs will be offered as part of our NVIDIA AI enterprise software platform for production deployment in the cloud or on-prem.

Nvidia’s GPUs that are meant for gaming on personal computers (PCs) can also be used for running generative AI applications on PCs; Nvidia and Microsoft has a partnership that help Windows to run LLMs up to 3x faster on PCs equipped with Nvidia’s GeForce RTX GPU

From the very start of our AI journey, we equipped GeForce RTX GPUs with CUDA Tensor cores. Now with over 100 million of an installed base, GeForce RTX GPUs are perfect for gamers, creators, AI enthusiasts, and offer unmatched performance for running generative AI applications on PCs. NVIDIA has full technology stack for deploying and running fast and efficient generative AI inference on GeForce RTX PCs…

…Yesterday, NVIDIA and Microsoft announced AI performance optimizations for Windows to help run LLMs up to 3x faster on NVIDIA GeForce RTX AI PCs.

Nvidia’s management is seeing game developers using the company’s AI services to create non-playable life-like characters in games

Top game developers, including NetEase Games, Tencent and Ubisoft are embracing NVIDIA Avatar Character Engine (sic) [ Avatar Cloud Engine ] to create lifelike avatars to transform interactions between gamers and non-playable characters.

Nvidia’s management thinks that the combination of generative AI and the Omniverse can drive the next wave of professional visualisation growth; the Ominverse has helped Wistron to reduce production cycle times by 50% and defect rates by 40%

We believe generative AI and Omniverse industrial digitalization will drive the next wave of professional visualization growth…

…Companies are using Omniverse to digitalize their workflows. Omniverse power digital twins enable Wistron, one of our manufacturing partners, to reduce end-to-end production cycle times by 50% and defect rates by 40%. 

Nvidia’s management sees generative AI driving a platform shift in the full computing stack

With generative AI, inference, which is now about fast token generation at massive scale, has become incredibly complex. Generative AI is driving a from-foundation-up full stack computing platform shift that will transform every computer interaction. From today’s information retrieval model, we are shifting to an answers and skills generation model of computing. AI will understand context and our intentions, be knowledgeable, reason, plan and perform tasks. We are fundamentally changing how computing works and what computers can do, from general purpose CPU to GPU accelerated computing, from instruction-driven software to intention-understanding models, from retrieving information to performing skills and, at the industrial level, from producing software to generating tokens, manufacturing digital intelligence.

Nvidia’s management sees token generation from LLMs driving multi-year build out of AI factories

Token generation will drive a multiyear build-out of AI factories…

… Large clusters like the ones built by Meta and Tesla are examples of the essential infrastructure for AI production, what we refer to as AI factories. These next-generation data centers host advanced full-stack accelerated computing platforms where the data comes in and intelligence comes out.

Nvidia’s management does not think that the demand they are seeing for the company’s AI chips is a pull-ahead of demand, because the the chips are being consumed

[Question] How are you ensuring that there is enough utilization of your products and that there isn’t a pull-ahead or a holding behavior because of tight supply, competition or other factors? 

[Answer] The demand for GPUs in all the data centers is incredible. We’re racing every single day. And the reason for that is because applications like ChatGPT and GPT-4o, and now it’s going to be multi-modality, Gemini and its ramp and Anthropic, and all of the work that’s being done at all the CSPs are consuming every GPU that’s out there. There’s also a long line of generative AI startups, some 15,000, 20,000 startups that are in all different fields, from multimedia to digital characters, of course, all kinds of design tool application, productivity applications, digital biology, the moving of the AV industry to video so that they can train end-to-end models to expand the operating domain of self-driving cars, the list is just quite extraordinary. We’re racing actually. Customers are putting a lot of pressure on us to deliver the systems and stand those up as quickly as possible. And of course, I haven’t even mentioned all of the sovereign AIs who would like to train all of their regional natural resource of their country, which is their data, to train their regional models. And there’s a lot of pressure to stand those systems up. So anyhow, the demand, I think, is really, really high and it outstrips our supply.

Nvidia’s management thinks that AI is not merely a chips problem – it is a system problem

The third reason has to do with the fact that we build AI factories. And this is becoming more apparent to people that AI is not a chip problem only. It starts, of course, with very good chips and we build a whole bunch of chips for our AI factories, but it’s a systems problem. In fact, even AI is now a systems problem. It’s not just one large language model. It’s a complex system of a whole bunch of large language models that are working together. And so the fact that NVIDIA builds this system causes us to optimize all of our chips to work together as a system, to be able to have software that operates as a system, and to be able to optimize across the system.

Nvidia’s management sees the highest performing AI chip as having the lowest total cost of ownership (TCO)

Today, performance matters in everything. This is at a time when the highest performance is also the lowest cost because the infrastructure cost of carrying all of these chips cost a lot of money. And it takes a lot of money to fund the data center, to operate the data center, the people that goes along with it, the power that goes along with it, the real estate that goes along with it, and all of it adds up. And so the highest performance is also the lowest TCO.

From the point of view of Nvidia’s management, customers do not mind buying Nvidia’s AI chips today even though better ones are going to come out tomorrow because they are still very early in their build-out of their AI infrastructure, and they want to ship AI advancements fast

[Question]  I’ve never seen the velocity that you guys are introducing new platforms at the same combination of the performance jumps that you’re getting…  it’s an amazing thing to watch but it also creates an interesting juxtaposition where the current generation of product that your customers are spending billions of dollars on is going to be not as competitive with your new stuff very, very much more quickly than the depreciation cycle of that product. So I’d like you to, if you wouldn’t mind, speak a little bit about how you’re seeing that situation evolve itself with customers. 

[Answer]  If you’re 5% into the build-out versus if you’re 95% into the build-out, you’re going to feel very differently. And because you’re only 5% into the build-out anyhow, you build as fast as you can… there’s going to be a whole bunch of chips coming at them, and they just got to keep on building and just, if you will, performance-average your way into it. So that’s the smart thing to do. They need to make money today. They want to save money today. And time is really, really valuable to them. Let me give you an example of time being really valuable, why this idea of standing up a data center instantaneously is so valuable and getting this thing called time-to-train is so valuable. The reason for that is because the next company who reaches the next major plateau gets to announce a groundbreaking AI. And the second one after that gets to announce something that’s 0.3% better. And so the question is, do you want to be repeatedly the company delivering groundbreaking AI or the company delivering 0.3% better?

All of Nvidia’s AI-related hardware products runs on its CUDA software; management thinks that AI performance for Nvidia AI-hardware users can improve over time simply from improvements that the company will be making to CUDA in the future

And all of it — the beautiful thing is all of it runs CUDA. And all of it runs our entire software stack. So if you invest today on our software stack, without doing anything at all, it’s just going to get faster and faster and faster. And if you invest in our architecture today, without doing anything, it will go to more and more clouds and more and more data centers and everything just runs. 

Shopify (NASDAQ: SHOP)

Shopify Magic is Shopify’s suite of AI products and management’s focus is on providing AI tools for merchants to simplify business operations and enhance productivity

Touching briefly on AI. Our unique position enables us to tap into the immense potential of AI for entrepreneurship and our merchants. Currently, the most practical applications of AI are found in tools that simplify business operations and enhance productivity, all of which we’ve been developing deeper capabilities with our AI product suite, Shopify Magic. 

Shopify’s management is using AI tools for precision marketing, and drove a 130% increase in merchant ads within its primary marketing channel from 2023 Q4 to 2024 Q1 while still being within payback guardrails

Our goal is to always get the most out of every existing channel up to our guardrail limits and continuingly find and experiment with new channels. That is what we build our tools and our AI models to do, and we’re using them to create some incredibly compelling opportunities. Let me give you a very recent example. At the end of last year and early into January, we drove significant efficiency improvements in one of our primary channels in performance marketing, where teams have created and leveraged advanced models using AI and machine learning, which now allows us to target our audiences with unprecedented precision. Using these models and strategies, we drove nearly 130% increase in merchant ads within our primary marketing channel from Q4 to Q1, while still remaining squarely within our payback guardrails.

Shopify has produced good revenue growth despite its headcount remaining flat for 3 quarters; management thinks Shopify can keep headcount growth low while the business continues to grow; the use of AI internally is an important element of how Shopify can continue to drive growth while keeping headcount growth low; an example of an internal use-case of AI is merchant support, where Shopify has (1) seen more than half of support interactions being assisted, and often fully-resolved, by AI, (2) been able to provide 24/7 live support in 8 additional languages that previously were offered only for certain hours, (3) decreased the duration of support interactions, (4) reduce the reluctance of merchants to ask questions, and (4) reduced the amount of toil on support staff

We know our team is one of our most valuable assets. And given that it makes up over half of our cost base, we believe we’ve architected ourselves to be faster and more agile, which has enabled us to consistently deliver 25% revenue growth, excluding logistics, all while keeping our headcount flat for 3 straight quarters. More importantly, because of the structure and the automation we have worked to put in place, we think we can continue to operate against very limited headcount growth while achieving a continued combination of consistent top line growth and profitability…

…We continue to remain disciplined on headcount with total headcount remaining essentially flat for the past 3 quarters, all while maintaining and, in fact, accelerating our product innovation capabilities and continuing the top line momentum of our business. How we leverage AI internally is an important element of how we are able to do that…

…During Q1, over half of our merchant support interactions were assisted with AI and often fully resolved with the help of AI. AI has enabled 24/7 live support in 8 additional languages that previously were offered only certain hours of the day. We have significantly enhanced the merchant experience. The average duration of support interactions has decreased. And the introduction of AI has helped reduce the reluctance that some merchants previously had towards asking questions that they might perceive as trivial or naive. Additionally, our support staff has experienced a significant reduction in the amount of toil that is part of their jobs. We are improving the merchant support process and achieving much greater efficiency than ever before.

Taiwan Semiconductor Manufacturing Company (NYSE: TSM)

TSMC’s management confirmed that there are no major damages to the company’s fabs and major operations from the recent earthquake in Taiwan – the largest in the region in 25 years – so there are no major disruptions to the supply of AI chips

On April 3, an earthquake of 7.2 magnitude struck Taiwan, and the maximum magnitude of our fabs was 5. Safety systems and protocols at our fabs were initiated immediately and all TSMC personnel are safe. Based on TSMC’s deep experience and capabilities in earthquake response and damage prevention as well as regular disasters trials, the overall tool recovery in our fabs reached more than 70% within the first 10 hours and were fully recovered by the end of the third day. There were no power outages, no structural damage to our fabs, and there’s no damage to our critical tools, including all our EUV lithography tools. That being said, a certain number of wafers in process were impacted and had to be scrapped, but we expect most of the lost production to be recovered in the second quarter and thus, minimum impact to our second quarter revenue. We expect the total impact from the earthquake to reduce our second quarter gross margin by about 50 basis points, mainly due to the losses associated with wafer scraps and material loss…

…Although it was largest earthquake in Taiwan in the last 25 years, we worked together tirelessly and were able to resume for operation at all our fab within 3 days with minimal disruptions, demonstrating the resilience of our operation in Taiwan.

TSMC’s management is seeing a strong surge in AI-related demand, and thinks that this supports their view of a structural acceleration in demand for energy-efficient computing

The continued surge in AI-related demand supports our already strong conviction that structural demand for energy-efficient computing is accelerating in an intelligent and connected world. 

TSMC’s management sees the company as a key enabler of AI; the increase in complexity of AI models, regardless of the approaches taken, requires increasingly powerful semiconductors, and this is where TSMC’s value increases, because the company excels at manufacturing the most advanced semiconductors

TSMC is a key enabler of AI applications. AI technology is evolving to use our increasingly complex AI models, which needs to be supported by more powerful semiconductor hardware. No matter what approach is taken, it requires use of the most advanced semiconductor process technologies. Thus, the value of our technology position is increasing as customers rely on TSMC to provide the most advanced process and packaging technology at scale, with a dependable and predictable cadence of technology offering. In summary, our technology leadership enable TSMC to win business and enables our customer to win business in the AI market.

TSMC’s management is seeing nearly every AI innovator working with the company

Almost all the AI innovators are working with TSMC to address the insatiable AI-related demand for energy-efficient computing power. 

TSMC’s management is forecasting the company’s revenue from AI processors to more than double in 2024 and account for low-teens percentage of total revenue; management expects AI processor revenue to grow at 50% annually over the next 5 years and account for more than 20% of TSMC’s total revenue by 2028; management has a narrow definition of AI processors and expect them to be the strongest growth driver for TSMC’s overall HPC (high performance computing) platform and overall revenue over the next few years

We forecast the revenue contribution from several AI processors to more than double this year and account for low teens percent of our total revenue in 2024. For the next 5 years, we forecast to grow at 50% CAGR and increase to higher than 20% of our revenue by 2028. Several AI processes are narrowly defined as GPUs, AI accelerators, and CPUs performing training and inference functions, and do not improve the networking, edge or on-device AI. We expect several AI processor to be the strongest driver of our HPC platform for growth and the largest contributor in terms of our overall incremental revenue growth in the next several years.

TSMC’s management thinks that strong HPC and AI demand means that it is strategically important for the company to expand its global manufacturing footprint

Given the strong HPC and AI-related demand, it is strategically important for TSMC to expand our global manufacturing footprint to continue to support our U.S. customers, increased customer trust, and expand our future growth potential.

TSMC has received strong support from the US government for its Arizona fabs and one of them has been upgraded to be a fab for 2nm process technology to support AI-demand, and it is scheduled for volume production in 2028; management is confident that the Arizona fabs will have the same quality as TSMC’s Taiwan fabs

In Arizona, we have received a strong commitment and support from our U.S. customers and plan to build 3 fabs, which help to create greater economies of scale..

…Our second fab has been upgraded to utilize 2-nanometer technologies to support a strong AI-related demand in addition to the previously announced 3-nanometer. We recently completed the taping of in which the last construction beam was raised into place and volume production is scheduled to begin in 2028…

…We are confident that once we begin volume production, we will be able to deliver the same level of manufacturing quality and reliability in each of our fab in Arizona as from our fab in Taiwan.

TSMC’s management believes the company’s 2nm technology is industry-leading and nearly every AI innovator is working with the company on its 2nm technology; management thinks 2nm will enable TSMC to capture AI-related growth opportunities in the years ahead

Finally, I will talk about our N2 status. Our N2 technology leads industry in addressing the industry’s insatiable need for energy-efficient computing, and almost all AI innovators are working with TSMC…

… With our strategy of continuous enhancement, N2 and its derivative will further extend our technology leadership position and enable TSMC to capture the AI-related growth opportunities well into future.

TSMC’s management is seeing very, very strong AI-related data center demand, while traditional server demand is slow; there is a shift in wallet-share from hyperscalers from traditional servers to AI servers and that is favourable for TSMC because TSMC has a lower presence in the traditional CPU-centric server space; TSMC is doubling its production capacity for AI-related data centre chips, but it’s still not enough to meet its customers’ demand

However, AI-related data center demand is very, very strong. And traditional server demand is slow, lukewarm…

…The budget for hyperscale player, their wallet-share shift from traditional server to AI server is favorable for TSMC. And we are able to capture most of the semiconductor content in an AI servers area as we defined GPU, ACA networking processor, et cetera. Well, we have a lower presence in those CPU-only, CPU-centric traditional server. So we expect our growth will be very healthy…

…Let me say it again, the demand is very, very strong, and we have done our best we put all the effort to increase the capacity. It probably more than double this year as compared with last year. However, not enough to meet the customers’ demand, and we leverage our OSAT partners that to complement of TSMC’s capacity to fulfill our customers need. Still not enough, of course. 

TSMC’s management is working on selling TSMC’s value in the manufacture of AI chips

[Question] I think it’s clear that AI is producing a large profit pool at your owners. And the HBM is also driving super normal returns for memory players. So my question is, does TSMC believe they’re getting their fair share of the returns in the AI value chain today? And is there a scope for TSMC to raise pricing for AI chips in the future?

[Answer] We always say that we want to sell our value, but it is a continuous process for TSMC. And let me tell you that we are working on it. We are happy that our customers are doing well. And if customers do well, TSMC does well.

TSMC’s management still expects the company’s capex intensity (capex as a percentage of revenue) to level off somewhere around the mid-30s range in the next several years even with the AI-boom, but they are ready to increase capex if necessary

[Question] My second question is just relating to the upward expectations you gave for the AI accelerators. Curious how that time, how you’re looking at the CapEx, if you say that we’re entering either higher growth or investment cycle, where capital intensity could need to rise up above that mid-30s range that you set 

[Answer] We work with our customers closely and our CapEx and capacity planning are always based on the long-term structural market demand profile that is underpinned by the multiyear megatrends….  The capital intensity, in the past few years, it was high as we invested heavily to meet the strong customer demand. Now the increase — the rate of increase for the capex is leveling off, so this year and the next several years, we are expecting that the capital intensity is somewhere at the mid-30s level. But as I just said, if there are opportunities in the future years, then we will invest accordingly.

TSMC’s management wants to support all of TSMC’s AI customers’ needs, and not just the needs of its major AI customer (presumably Nvidia)

 We want to make sure that all our customers get supported, probably not enough this year. But for next year, we try. We try very hard. And you mentioned about giving up some market share, that’s not my consideration. My consideration is to help our customers to be successful in their market…

…[Question] So since your major customers said there’s no room for other type of AI computing chips, but it seems like TSMC is happy to assist some similar customers, right? So is that right interpretation about your comments.

[Answer] Yes.

Most of TSMC’s AI customers are using the 5nm or 4nm technologies, but they are working with TSMC on even more advanced nodes – such as 3nm and 2nm – because the advanced nodes are more energy-efficient, and energy efficiency in AI data centres is really important; in the past, TSMC’s then-leading edge chips only see smartphone demand, but with 2nm, TSMC will see demand from smartphones and HPC, so the early-revenue from 2nm is expected to be even larger than 3nm’s early-revenue 

[Question] I think currently, most of the AI accelerator, mostly in 5-nanometers, which is N minus 1 comparing to a smartphone for now. So when do we expect them to catch up or surpass in terms of technology node? Do we see them to be the technology driver in 2 nanometers or above?

[Answer] Today, all the AI accelerators, most of them are in the 5- or 4-nanometer technology. My customers are working with TSMC for the next node, even for the next, next node, they have to move fast because, as I said, the power consumption has to be considered in the AI data center. So the energy-efficient is very important. So our 3-nanometer is much better than the 5-nanometer. And again, it will be improved in the 2-nanometer. So all I can say is all my customers are working on this kind of trend from 4-nanometer to 3 to 2…

…[Question] Do we see a bigger revenue in the first 2 years of the 2 nanometers because in the past, it’s only smartphone, but in 2-nanometer, it would be both smartphone and HPC customers.

[Answer] With the demand that we’re seeing, we do expect N2 revenue contribution to be even larger than N3, just like 3 is a larger contribution or larger node than 5, et cetera, et cetera.

TSMC’s management is seeing die sizes increase with edge-AI or on-device AI; management thinks that the replacement cycle for smartphones and PCs will be a little accelerated in the future and the edge-AI trend will be very positive for TSMC

Let me mention the edge-AI or the on-device AI, the first order of magnitude is the die size. We saw with AI for neuro processor inside, the die size will be increased, okay? That’s the first we observed. And it’s happening. And then for the future, I would think that replacement cycle for smartphone and kind of a PC will be accelerated a little bit in the future, at least. It’s not happening yet, but we do expect that will happen soon. And all in all, I would say that on-device AI will be very positive for TSMC because we kept the larger share of the market. 

Tencent (NASDAQ: TCEHY)

Engagement of Weixin users is increasingly supplemented by consumption of content in chat at moments and recommended content on video accounts and mini programs; this was driven by AI recommendations 

For Weixin, users are increasingly supplementing their stable consumption of social graph supply content in chat at moments with consumption of algorithmically recommended content in official accounts and video accounts and engagement with Mini Programs diverse range of services.  This trend benefits from our heavy investment in AI, which makes the recommendation better and better over time.

Official accounts achieved healthy year-on-year pageview growth, driven AI-powered recommendation algorithms

For official accounts, which enable creators to share text and images and chosen topics with interested followers, it achieved healthy year-on-year pageview growth. As AI-powered recommendation algorithms allow us to provide targeted high-quality content more effectively.

Tencent’s online advertising revenue was up 26% in 2024 Q1 because of increased engagements from AI-powered ad targeting; ad spend from all major categories increased in 2024 Q1 except for automotives; during the quarter, Tencent upgraded its ad tech platform and made generative AI-powered ad creation tools available to boost ad creation efficiency and better targeting

For online advertising, our revenue was RMB 26.5 billion in the quarter up 26% year-on-year, benefiting from increased engagements in AI-powered ad targeting. Ad spend from all major categories except automotive increased year-on-year, particularly from games, internet services and consumer goods sectors. During the quarter, we upgraded our ad tech platform to help advertisers manage ad campaigns more effectively, and we made generative AI-powered ad creation tools available to all advertisers. These initiatives enable advertisers to create ads more efficiently and to deliver better targeting.

Hunyuan (Tencent’s foundational LLM) was scaled up using the mixture of experts approach; management is deploying Hunyuan in more Tencent services; management is open-sourcing a version of Hunyuan that provides text-image generative AI

And for Hunyuan, the main model achieved significant progress as we’ve scaled up using the mixture of experts approach, and we’re deploying Hunyuan in more of our services. Today, we announced that we’re making a version of Hunyuan providing text image generative AI available on an open source basis.

Tencent’s operating capex was RMB6.6b in 2024 Q1, up massively from a low base in 2023 Q1 but down slightly sequentially, because of spending on GPUs and servers to support Hunyuan and the AI ad recommendation algo

Operating CapEx was RMB 6.6 billion, up 557% year-on-year from a low base quarter last year, mainly driven by investment in GPUs and servers to support our Hunyuan and AI ad recommendation algorithms.

Tencent’s management expects advertising revenue growth to decelerate from 2024 Q1’s level, but still expects to outpace the broader industry because (1) Tencent’s ad load is still small relative to the advertising real estate available, and (2) AI will help the advertising business and can easily double or even triple Tencent’s currently low click-through rates; management thinks Tencent’s advertising business will benefit from AI disproportionately vis-a-vis competitors who also use AI because Tencent has been under-monetising and has lower click-through rates, so any AI-driven improvements will have a bigger impact; Hunyuan is part of the AI technologies that management has deployed for the advertising business

Around advertising, I’d say that, as you would expect, given the economies mix, advertiser sentiment is also quite mixed and it’s certainly a challenging environment in which to set advertising. The first quarter for us is a slightly unusual quarter because it’s a small quarter for advertising due to the Chinese New Year effect. And so sometimes the accelerations or the decelerations get magnified as a result. So we would expect our advertising growth to be less rapid in subsequent quarters of the year than it was in the first quarter and more similar to consensus expectations for our advertising revenue growth for the rest of the year. But that said, we think that we are in a good position to continue taking share of the market at a rapid rate, given we’re very early in increasing our ad load on video accounts, which is currently around 1/4 of the ad loads of our major competitors with short video products. 

And also given we’re early in capturing the benefits of deploying AI to our ad tech stack. And we think that we will — we are benefiting and will continue to benefit disproportionately from applying AI to our ad tech because historically, as a social media platform, our click-through rates were low. And so starting from that lower base, we can — we have seen we can double or triple click-through rates in a way that’s not possible for ad services that are starting from much higher click through rates…

… [Question] In the future, do you think like under the AI developments like our competitors such as like ByteDance or Alibaba, they also applies AI to their ad business so how do you think that AI will drive to add market share to change in the longer term?

[Answer] Your question around a number of competitors are obviously applying AI as well. And we believe that all of them will benefit from AI, too. But we think that the biggest beneficiaries will be those companies, of which we are one that have very substantial under monetized time spent and now able to monetize that time spend more effectively by deploying AI because the deployment of AI enables an upward structural shift in click-through rates, and that shift is most pronounced for those inventories where the click-through rates were lower to begin with, such as the social media inventory. Those tools also allow advertisers who previously were able to create advertisements for search, which are text in nature, but not to create advertisements for social media, which are image and video in nature, to now use generative AI to create advertisements to social media. So in general, we think there’ll be a reallocation of advertising spend toward those services, which have high time spent, high engagement and are now able to deliver increasing click through rates, increasing transaction volume more commensurate with the time spent and engagement superiority…

…  So on ad tech, we’re innovating around the process of targeting the ads using artificial intelligence. We’re innovating around helping advertisers manage their advertising campaigns. And then most recently, we’ve been — we are now deploying Hunyuan to facilitate advertisers, creating the advertising content.

Tencent’s management thinks that WeChat will be a great distribution channel for AI products, but they are still figuring out the best use case for AI (including Tencent’s own Hunyuan LLM); management is actively testing, and they will roll out the products they think are the best over time

I think we do believe that with the right product than our WeChat platform and our other products, which have a lot of user engagement would be great — will be great distribution channels for these AI products. But I think at this point in time, everybody is actually trying out different products that may work. No one has really come up with a killer application yet with the exception of probably OpenAI, that question and answer from it so I think you should be confident that we have been developing the technology, and we are having a best-in-class technology in Hunyuan and at the same time, we are actively creating and testing out different products to see what would make sense for our existing products and as the time comes, these products will be rolled out on our platform.

Tencent’s management thinks that Hunyuan is currently best being deployed in Tencent’s gaming business for customer service purposes; management has been deploying AI in Tencent’s games, but not necessarily generative AI; Hunyuan will be useful for developing games when it gains multi-modal capabilities, especially in creating high-quality videos, but it will be some time before Hunyuan reaches that level

I think for Hunyuan — it can be assisting game business in multiple ways. Right now, the best the best contributor is actually on the customer service front. When Hunyuan is actually deployed to answer questions and the customer service bought for a lot of our games is actually achieving very high customer satisfaction level. And AI, in general, has already been deployed in our games, but not necessarily the generative AI technology yet. In terms of Hunyuan and, I think, over time, when we actually sort of can move Hunyuan into a multi-modal and especially if we can start creating really high-quality, high fidelity videos, then that would actually be helpful. Before that happens, Hunyuan can actually sort of be using MPCs and create a certain sort of interactive experiences but it’s not going to be able to take over the very heavy growth of content creation in gaming yet. I think you’ll probably be a couple more generations before it can be for game production.

Tesla (NASDAQ: TSLA)

Tesla’s FSD v12 is a pure AI-based self driving technology; FSD v12 is now turned on for all North American Tesla vehicles – around 1.8 million vehicles – that are running on Hardware 3 or later and it is used on around half of the vehicles, with the percentage of users increasing each week; more than 300 billion miles have been driven with FSD v12; management thinks that it’s only a matter of time before Tesla’s autonomous driving capabilities exceeds human-reliability

Regarding FSD V12, which is the pure AI-based self-driving, if you haven’t experienced this, I strongly urge you to try it out. It’s profound and the rate of improvement is rapid. And we’ve now turned that on for all cars, with the cameras and inference computer, everything from Hardware 3 on, in North America. So it’s been pushed out to, I think, around 1.8 million vehicles, and we’re seeing about half of people use it so far and that percentage is increasing with each passing week. So we now have over 300 billion miles that have been driven with FSD V12…

…I think it should be obvious to anyone who’s driving V12 in a Tesla that it is only a matter of time before we exceed the reliability of humans and we’ve not much time with that. 

Tesla’s management believes that the company’s vision-based approach with end-to-end neural networks for full self driving is better than other approaches, because it mimics the way humans drive, and the global road networks are designed for biological neural nets and eyes

Since the launch of Full Self-Driving — Supervised Full Self-Driving, it’s become very clear that the vision-based approach with end-to-end neural networks is the right solution for scalable autonomy. And it’s really how humans drive. Our entire road network is designed for biological neural nets and eyes. So naturally, cameras and digital neural nets are the solution to our current road system…

… I think we just need to — it just needs to be obvious that our approach is the right approach. And I think it is. I think now with 12.3, if you just have the car drive you around, it is obvious that our solution with a relatively low-cost inference computer and standard cameras can achieve self-driving. No LiDARs, no radars, no ultrasonic, nothing.

Tesla has reduced the subscription price of FSD to US$99 a month; management is talking to one major auto manufacturer on licensing Tesla’s FSD software; it will take time for third-party automakers to use Tesla’s autonomous driving technology as a massive design change is needed for the vehicles even though all that is needed is for cameras and an inference computer to be installed

To make it more accessible, we’ve reduced the subscription price to $99 a month, so it’s easy to try out…

…We’re in conversations with one major automaker regarding licensing FSD…

…I think we just need to — it just needs to be obvious that our approach is the right approach. And I think it is. I think now with 12.3, if you just have the car drive you around, it is obvious that our solution with a relatively low-cost inference computer and standard cameras can achieve self-driving. No LiDARs, no radars, no ultrasonic, nothing… No heavy integration work for vehicle manufacturers…

… So I wouldn’t be surprised if we do sign a deal. I think we have a good chance we do sign a deal this year, maybe more than one. But yes, it would be probably 3 years before it’s integrated with a car, even though all you need is cameras and our inference computer. So just talking about a massive design change.

Tesla’s management has been expanding the company’s core AI infrastructure and the company is no longer training-constrained; Tesla has 35,000 H100 GPUs that are currently working, and management expects to have 85,000 H100 GPUs by end-2024 for AI training

Over the past few months, we’ve been actively working on expanding Tesla’s core AI infrastructure. For a while there, we were training-constrained in our progress. We are, at this point, no longer training-constrained, and so we’re making rapid progress. We’ve installed and commissioned, meaning they’re actually working, 35,000 H100 computers or GPUs. GPU is a wrong word, they need a new word. I always feel like a [ wentz ] when I say GPU because it’s not. GPU stands — G stands for graphics. Roughly 35,000 H100S are active, and we expect that to be probably 85,000 or thereabouts by the end of this year in training, just for training. 

Tesla’s AI robot, Optimus, is able to do simple factory tasks and management thinks it can do useful tasks by the end of this year; management thinks Tesla can sell Optimus by the end of next year; management still thinks that Optimus will be an incredibly valuable product if it comes to fruition; management thinks that Tesla is the best-positioned manufacturer of humanoid robots with efficient AI inference to be able to reach production at scale

[Question] What is the current status of Optimus? Are they currently performing any factory tasks? When do you expect to start mass production?

[Answer] We are able to do simple factory tasks or at least, I should say, factory tasks in the lab. In terms of actually — we do think we will have Optimus in limited production in the factory — in natural factory itself, doing useful tasks before the end of this year. And then I think we may be able to sell it externally by the end of next year. These are just guesses. As I’ve said before, I think Optimus will be more valuable than everything else combined. Because if you’ve got a sentient humanoid robots that is able to navigate reality and do tasks at request, there is no meaningful limit to the size of the economy. So that’s what’s going to happen. And I think Tesla is best positioned of any humanoid robot maker to be able to reach volume production with efficient inference on the robot itself.

The vision of Tesla’s management for autonomous vehicles is for the company to own and operate some autonomous vehicles within a Tesla fleet, and for the company to be an Airbnb- or Uber-like platform for other third-party owners to put their vehicles into the fleet; management thinks Tesla’s fleet can be tens of millions of cars worldwide – even more than 100 million – and as the fleet grows, it will act as a positive flywheel for Tesla in terms of producing data for training

And something I should clarify is that Tesla will be operating the fleet. So you can think of like how Tesla — you think of Tesla like some combination of Airbnb and Uber, meaning that there will be some number of cars that Tesla owns itself and operates in the fleet. There will be some number of cars — and then there’ll be a bunch of cars where they’re owned by the end user. That end user can add or subtract their car to the fleet whenever they want, and they can decide if they want to only let the car be used by friends and family or only by 5-star users or by anyone. At any time, they could have the car come back to them and be exclusively theirs, like an Airbnb. You could rent out your guestroom or not any time you want. 

So as our fleet grows, we have 7 million cars — 9 million cars, going to eventually tens of millions of cars worldwide. With a constant feedback loop, every time something goes wrong, that gets added to the training data and you get this training flywheel happening in the same way that Google Search has the sort of flywheel. It’s very difficult to compete with Google because people are constantly doing searches and clicking and Google is getting that feedback loop. So same with Tesla, but at a scale that is maybe difficult to comprehend. But ultimately, it will be tens of millions…

… And then I mean if you get like to the 100 million vehicle level, which I think we will, at some point, get to, then — and you’ve got a kilowatt of useable compute and maybe your own Hardware 6 or 7 by that time, then you really — I think you could have on the order of 100 gigawatts of useful compute, which might be more than anyone more than any company, probably more than any company.

Tesla’s management thinks that the company can sell AI inference compute capacity that’s sitting in Tesla vehicles when they are not in use; Tesla cars are running Hardware 3 and Hardware 4 now, while Hardware 5 is coming; unlike smartphones or computers, the computing capacity of Tesla vehicles is entirely within Tesla’s control, and the company has skills on deploying compute workloads to each individual vehicle

I think there’s also some potential here for an AWS element down the road where if we’ve got very powerful inference because we’ve got a Hardware 3 in the cars, but now all cars are being made with Hardware 4. Hardware 5 is pretty much designed and should be in cars hopefully towards the end of next year. And there’s a potential to run — when the car is not moving, to actually run distributed inference. So kind of like AWS, but distributed inference. Like it takes a lot of computers to train an AI model, but many orders of magnitude less compute to run it. So if you can imagine a future [ path ] where there’s a fleet of 100 million Teslas, and on average, they’ve got like maybe a kilowatt of inference compute, that’s 100 gigawatts of inference compute distributed all around the world. It’s pretty hard to put together 100 gigawatts of AI compute. And even in an autonomous future where the car is perhaps used instead of being used 10 hours a week, it is used 50 hours a week. That still leaves over 100 hours a week where the car inference computer could be doing something else. And it seems like it will be a waste not to use it…

…And then I mean if you get like to the 100 million vehicle level, which I think we will, at some point, get to, then — and you’ve got a kilowatt of useable compute and maybe your own Hardware 6 or 7 by that time, then you really — I think you could have on the order of 100 gigawatts of useful compute, which might be more than anyone more than any company, probably more than any company…

…Yes, probably because it takes a lot of intelligence to drive the car anyway. And when it’s not driving the car, you just put this intelligence to other uses, solving scientific problems or answer in terms of [ this horse ] or something else… We’ve already learned about deploying workloads to these nodes… And unlike laptops and our cell phones, it is totally under Tesla’s control. So it’s easier to see the road products plus different nodes as opposed to asking users for permission on their own cell phones would be very tedious… 

… So like technically, yes, I suppose like Apple would have the most amount of distributed compute, but you can’t use it because you can’t get the — you can’t just run the phone at full power and drain the battery. So whereas for the car, even if you’re a kilowatt-level inference computer, which is crazy power compared to a phone, if you’ve got 50 or 60 kilowatt hour pack, it’s still not a big deal. Whether you plug it or not, you could run for 10 hours and use 10 kilowatt hours of your kilowatt of compute power.

Safety is very important for Tesla; management has been conducting safety-training for Tesla’s AI-powered self driving technology through the use of millions of clips of critical safety events collected from Tesla vehicles; the company runs simulations for safety purposes before pushing out a new software version to early users and before it gets pushed to external users; once the new software is with external users, it’s constantly monitored by Tesla; FSD v12’s feedback loop of issues, fixes, and evaluations happens automatically because the AI model learns on its own based on data it is getting

Yes, we have multiple years of validating the safety. In any given week, we train hundreds of neural networks that can produce different trajectories for how to drive the car, replay them through the millions of clips that we have already collected from our users and our own QA. Those are like critical events, like someone jumping out in front or like other critical events that we have gathered database over many, many years, and we replay through all of them to make sure that we are net improving safety. 

And then we have simulation systems. We also try to recreate this and test this in close to fashion. And some of this is validated, we give it to our QA networks. We have hundreds of them in different cities, in San Francisco, Los Angeles, Austin, New York, a lot of different locations. They are also driving this and collecting real-world miles, and we have an estimate of what are the critical events, are they net improvement compared to the previous week builds. And once we have confidence that the build is a net improvement, then we start shipping to early users, like 2,000 employees initially that they would like it to build. They will give feedback on like if it’s an improvement or they’re noting some new issues that we did not capture in our own QA process. And only after all of this is validated, then we go to external customers.

And even when we go external, we have like live dashboards of monitoring every critical event that’s happening in the fleet sorted by the criticality of it. So we are having a constant pulse on the build quality and the safety improvement along the way. And then any failures like Elon alluded to, we’ll get the data back, add it to the training and that improves the model in the next cycle. So we have this like constant feedback loop of issues, fixes, evaluations and then rinse and repeat.

And especially with the new V12 architecture, all of this is automatically improving without requiring much engineering interventions in the sense that engineers don’t have to be creative and like how they code the algorithms. It’s mostly learning on its own based on data. So you see that, okay, every failure or like this is how a person chooses, this is how you drive this intersection or something like that, they get the data back. We add it to the neural network, and it learns from that trained data automatically instead of some engineers saying that, oh, here, you must rotate the steering wheel by this much or something like that. There’s no hard inference conditions. If everything is neural network, it’s pretty soft, it’s probabilistic and circular. That’s probabilistic distribution based on the new data that it’s getting.

Tesla’s management has good insight on the level of improvement Tesla’s AI-powered self-driving technology can be over a 3-4 month time frame, based on a combination of model size scaling, data scaling, training compute scaling, and architecture scaling

And we do have some insight into how good the things will be in like, let’s say, 3 or 4 months because we have advanced models that our far more capable than what is in the car, but have some issues with them that we need to fix. So they are there’ll be a step change improvement in the capabilities of the car, but it will have some quirks that are — that need to be addressed in order to release it. As Ashok was saying, we have to be very careful in what we release the fleet or to customers in general. So like — if we look at say 12.4 and 12.5, which are really could arguably even be V13, V14 because it’s pretty close to a total retrain of the neural nets and in each case, are substantially different. So we have good insight into where the model is, how well the car will perform, in, say, 3 or 4 months…

… In terms of scaling, people in here coming and they generally talk about models scaling, where they increase the model size a lot and then their corresponding gains in performance, but we have also figured out scaling loss and other access in addition to the model side scaling, making also data scaling. You can increase the amount of data you use to train the neural network and that also gives similar gains and you can also scale up by training compute. You can train it for much longer and one more GPUs or more Dojo nodes, and that also gives better performance. And you can also have architecture scaling where you count with better architectures for the same amount of compute produce better results. So a combination of model size scaling, data scaling, training compute scaling and the architecture scaling, we can basically extrapolate, okay, with the continue scaling based at this ratio, we can predict future performance. 

The Trade Desk (NASDAQ: TSLA)

Trade Desk’s management will soon roll out a game-changing AI-fueled forecasting tool on the company’s Kokai platform

We are quickly approaching some of the biggest UX and product rollouts of Kokai that nearly all of our customers will begin to use and see benefits from over the next few quarters, including a game-changing AI-fueled forecasting tool.

Trade Desk’s management has been using AI since 2016; management has always thought about AI as a copilot for humans even before Trade Desk was founded

We’ve been deploying AI in our platform since we launched Koa in 2016…

… To that end, we’ve known since before our company existed that the complexity of assessing millions of ad opportunities every second, along with hundreds of variables for each impression, is beyond the scope of any individual human. We have always thought about AI as a copilot for our hands-on keyboard traders.

Through Kokai, Trade Desk is bringing AI to many decision-points in the digital advertising process; Trade Desk is also incorporating AI into new relevance indices in Kokai for advertisers to better understand the relevance of different ad impressions in reaching their target audience; US Cellular used Trade Desk’s TV Quality Index to improve its conversion rate by 71%, reach 66% more households, and decrease cost per acquisition by 24%

And with Kokai, we are bringing the power of AI to a broader range of key decision points than ever, whether it’s in relevant scoring forecasting, budget optimization, frequency management or upgraded measurement. AI is also incorporated into a series of new indices that score relevance, which advertisers can use to better understand the relevance of different ad impressions in reaching their target audience. For example, U.S. Cellular worked with their agency, Harmelin Media, to leverage our TV Quality Index to better reach new customers. Their conversion rates improved 71%. They reached 66% more households by optimizing frequency management, and their cost per acquisition decreased 24%. I think it’s important to understand how we’re putting AI to work in Kokai because this kind of tech dislocation will bring new innovators. 

Visa (NYSE: V)

Visa’s management is using AI to improve the company’s risk offerings; the company’s Visa Protect for account-to-account payments feature is powered by AI-based fraud detection models; another of the features, Visa Deep Authorization, is powered by a deep-learning recurrent neural network model for risk scoring of e-commerce payments specifically in the USA

Across our risk offerings, we continue to bolster them through our technology, innovation, and AI expertise and are expanding their utility beyond the Visa network. Recently, we announced 3 such capabilities in our Visa Protect offering. The first is the expansion of our signature solutions, Visa Advanced Authorization and Visa Risk Manager for non-Visa card payments, making them network-agnostic. This allows issuers to simplify their fraud operations into a single fraud detection solution. The second is the release of Visa Protect for account-to-account payments, our first fraud prevention solution built specifically for real-time payments, including P2P digital wallets, account-to-account transactions and Central Bank’s instant payment systems. Powered by AI-based fraud detection models, this new service provides a real-time risk score that can be used to identify fraud on account-to-account payments. We’ve been piloting both of these in a number of countries, and our strong results thus far have informed our decision to roll these out globally. The third solution is Visa Deep Authorization. It is a new transaction risk scoring solution tailored specifically to the U.S. market to better manage e-commerce payments powered by a world-class deep-learning recurrent neural network model and petabytes of contextual data…

…What we found in the U.S. e-commerce market is that, on the one hand, it’s the most developed e-commerce market on the planet. On the other hand, it’s become the place of the most sophisticated fraud and attack vectors that we see anywhere in the world. And so what we are bringing to market with Visa Deep Authorization is an e-commerce transaction risk scoring platform and capability that is specifically tailored and built for the unique sets of attack vectors that we’re seeing in the U.S. So as I was mentioning in my prepared remarks, it’s built on deep learning technology that’s specifically tuned to some of the sequential and contextual view of accounts that we’ve had in the U.S. market. 

Wix (NASDAQ: WIX)

Wix’s management released its AI website builder in 2024 Q1, which is the company’s cornerstone product; the AI website builder utilises a conversational AI chat experience where users describe their intent and goals, and it is based on Wix’s decade-plus of knowledge in website creation and user behaviour; the AI-generated sites include all relevant pages, business solutions (such as scheduling and e-commerce), and functions; management thinks the AI website builder is a unique product in the market; management is seeing strong utilisation of the AI website builder, with hundreds of thousands of sites already been created in a few months since launch by both Self Creators and Partners

Notably, this quarter, we released the highly anticipated AI website builder. This is our cornerstone AI product. It leverages our 10-plus years of web creation expertise and unparalleled knowledge based on users’ behavior through a conversational AI chat experience. Users describe their intent and goals. Our AI technology then creates a professional, unique, and fully built-out website that meets the users’ needs. Importantly, the AI-generated site includes all relevant pages with personalized layout themes, text, images and business solutions such as scheduling, e-commerce and more. Best of all, this website are fully optimized with Wix-reliable infrastructure, including security and performance as well as built in marketing, SEO, CRM and analytics tools. There is truly nothing like this on the market. Excitingly, feedback on the AI website building has been incredible. In just a few short months since its launch, hundreds of thousands of sites have been already been created using this tool by both Self Creators and Partner. This strong response and utilization is a testament to the depth of our AI expertise and strength of our product. 

Wix released AI-powered image enhancement tools within Wix Product Studio in April which allow users to edit images in a high-quality manner through prompts

In April, we released a suite of AI-powered image enhancement tools that provide users with the capability to create professional images on their own. High-quality images are an essential part of a professional website but often hard to achieve without the help of professional photographer. New users will be able to easily erase objects, generate images, edit them to add or replace objects with a simple prompt, all without ever leaving the Wix Product Studio. 

Wix will be releasing more AI products in 2024; the upcoming products include AI business assistants; the AI business assistants are in beta testing and management is seeing great feedback

This new capabilities are just the start of a robust pipeline of AI-enabled products still to come this year, including a variety of vertical AI business assistants that will be released for the year. A couple of these assistants are currently in beta testing and seeing great results and feedback. 

Wix is seeing that its AI products are resulting in better conversion of users into premium subscribers; management believes that Wix’s AI products will be a significant driver of Self Creators growth in the years ahead

We are seeing a tangible benefit from our entire AI offering particularly a better conversions among users into premium subscription. I strongly believe that our AI capability will be significant — a significant driver of Self Creators growth in 2024 and beyond.

 Wix’s AI tools will be exposed very frequently to both existing and new users of the Wix platform

[Question] I wanted to kind of follow on to that and just kind of understand with respect to the AI tools. Do you see this primarily impacting the new customers? 

[Answer] When users are building their websites, all the website creation tools are visible to them and are helping them. Most of our users will stay a few years or more than that with the same website and sometimes — and they’ll update it, but they’re not going to recreate it. So, in that term, of course, the exposure is limited. But the integration of the vertical assistance is something that means that every time you go to the website, you’re going to have a recommendation, and the ideas and things you can do with AI. So, the exposure will be pretty much every time you go into the website. And that is significantly higher. And if you think about the fact that we have a lot of people that run their business in top of Wix, it means that all of those guys will be daily or almost daily exposed to new products with AI…

…You’re going to find AI tools, but they are not going to replace what you already know how to do. Sometimes, if you want to change an image, for example, it’s easier to click on change image instead of writing to the prompt, hey, please change the third image from the top, right? So, it’s always about the combination of how you do things in a balanced way, while allowing users to feel comfortable with the changes, not move beyond that. 

Wix’s management believes that AI will be a boom for new technologies and innovation and will lead to more growth for Wix

I believe that there’s so much potential for new things coming with AI, so much potential with new things coming with market trends and new technologies introduced into the market that I believe that we’re going to continue to see significant innovation, growing innovation coming from small businesses and bigger businesses in the world, which will probably result in the formation of additional growth for us. 

Zoom Video Communications (NASDAQ: ZM)

Zoom is now far beyond just video conferencing, and AI is infused across its platform

Our rapid innovation over the years has taken us far beyond video conferencing. Every step of the way has been guided by our mission to solve customer problems and enable greater productivity. In the process, we have very deliberately created a communication and collaboration powerhouse with AI infused natively across the platform.

Zoom’s management announced Zoom Workplace, an AI-powered collaboration platform in March; Zoom Workplace already has AI-powered features but will soon have Ask AI Companion; Zoom Workplace also improves other Zoom products through AI Companion capabilities; the AI features in Zoom Workplace are provided at no additional cost

In March we announced Zoom Workplace, our AI-powered collaboration platform designed to help our customers streamline communications, improve productivity, increase employee engagement, and optimize in-person time. Within the launch of Zoom Workplace are new enhancements and capabilities like multi-speaker view, document collaboration, AI-powered portrait lighting, along with upcoming features and products like Ask AI Companion, which will work across the platform to help employees make the most of their time. The Workplace launch also boosts Zoom Phone, Team Chat, Events and Whiteboard with many more AI Companion capabilities to help make customers more productive…

…When you look at our Workplace customers, guess what, AI is not only a part of that but also at no additional cost, right? So that is our vision.

Expedia has signed a quadruple-digit seat deal for Zoom Revenue Accelerator, which includes AI products that can help Expedia to drive revenue

Let me thank Expedia, who needs no introduction, for becoming a Lighthouse Zoom Revenue Accelerator customer in the quarter, leaning heavily into our AI products to drive revenue. A power user of Zoom Phone for years, they wanted to better automate workflows, coach sellers and drive efficiencies. We partnered with them on an initial quadruple-digit seat Zoom Revenue Accelerator deal, which includes working directly with their team to improve and tailor the product based on their business model and industry-specific use case.

Centerstone, a nonprofit organisation, expanded Zoom Phone and Zoom Contact Center in 2024 Q1 to leverage AI to provide better care for its beneficiaries

Let me also thank Centerstone, a nonprofit health system specializing in mental health and substance use disorder treatments for individuals, families, and veterans, for doubling down on Zoom. Seeing strong value from their existing Zoom Meetings, Phone and Rooms deployment, in Q1, they expanded Zoom Phone and added Zoom Contact Center in order to leverage AI to provide better care, and Zoom Team Chat in order to streamline communications all from a single platform.

Zoom AI Companion is now enabled in >700,000 customer accounts just 8 months after launch; AI Companion improves the value proposition of all of Zoom’s products and it’s provided to customers without charging customers more; AI Companion also helps Zoom improve monetisation because its presence in Zoom’s Business Services enables Zoom to charge a premium price because the AI features are a key differentiator; management will leverage AI Companion to build a lot of new things

Zoom AI Companion has grown significantly in just eight months with over 700,000 customer accounts enabled as of today. These customers range all the way from solopreneurs up to enterprises with over 100,000 users…

… I think AI Companion not only help our Meetings, Phone, or Team Chat, it’s across the entire Zoom Workplace platform plus all the Business Services, right? Our approach, if you look at our Workplace, the deployment, right, for the entire collaboration platform not only makes all those services better but also customers appreciate it, right, without charging the customers more, right? We do add more value to customers at no additional cost, right? That’s kind of the power part of the Zoom company. At the same time, in terms of monetization, as I mentioned earlier, if you look at our Business Services, AI is a key differentiation, right, AI and we charge a premium price as well, and that’s the value. At the same time, we also are going to leverage AI Companion to build a lot of new things, new services like Ask AI that will be introduced later this year and also some other new services that we’re working on as well.

One of Zoom’s management’s key priorities is to embed AI across all of Zoom Workplace and Business Services

Embedding AI across all aspects of Zoom Workplace and Business Services is a key priority as we continue to drive productivity and engagement for our customers.

---

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Adobe, Alphabet, Amazon, Apple, Coupang, Datadog, Etsy, Fiverr, Mastercard, Meta Platforms, Microsoft, Netflix, Shopify, TSMC, Tesla, The Trade Desk, Visa, Wix, and Zoom. Holdings are subject to change at any time.